temporal-twins-anon commited on
Commit
a24cb8f
·
verified ·
1 Parent(s): a6a52dc

Add Temporal Twins benchmark release v0.1

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. DATASET_CARD.md +458 -0
  2. LICENSE +203 -0
  3. LICENSE-DATA +18 -0
  4. MANIFEST.sha256 +126 -0
  5. README.md +96 -0
  6. README_REPO.md +437 -0
  7. RELEASE_CHECKLIST.md +22 -0
  8. configs/default.yaml +29 -0
  9. configs/paper_suite_reference.yaml +25 -0
  10. configs/temporal_twins_calib.yaml +29 -0
  11. croissant.json +796 -0
  12. data/.DS_Store +0 -0
  13. data/README_GENERATION.md +120 -0
  14. data/_export_summary.csv +21 -0
  15. data/easy/.DS_Store +0 -0
  16. data/easy/seed_0/audit_summary.csv +2 -0
  17. data/easy/seed_0/config.yaml +30 -0
  18. data/easy/seed_0/matched_pairs.parquet +3 -0
  19. data/easy/seed_0/schema.json +84 -0
  20. data/easy/seed_0/transactions.parquet +3 -0
  21. data/easy/seed_1/audit_summary.csv +2 -0
  22. data/easy/seed_1/config.yaml +30 -0
  23. data/easy/seed_1/matched_pairs.parquet +3 -0
  24. data/easy/seed_1/schema.json +84 -0
  25. data/easy/seed_1/transactions.parquet +3 -0
  26. data/easy/seed_2/audit_summary.csv +2 -0
  27. data/easy/seed_2/config.yaml +30 -0
  28. data/easy/seed_2/matched_pairs.parquet +3 -0
  29. data/easy/seed_2/schema.json +84 -0
  30. data/easy/seed_2/transactions.parquet +3 -0
  31. data/easy/seed_3/audit_summary.csv +2 -0
  32. data/easy/seed_3/config.yaml +30 -0
  33. data/easy/seed_3/matched_pairs.parquet +3 -0
  34. data/easy/seed_3/schema.json +84 -0
  35. data/easy/seed_3/transactions.parquet +3 -0
  36. data/easy/seed_4/audit_summary.csv +2 -0
  37. data/easy/seed_4/config.yaml +30 -0
  38. data/easy/seed_4/matched_pairs.parquet +3 -0
  39. data/easy/seed_4/schema.json +84 -0
  40. data/easy/seed_4/transactions.parquet +3 -0
  41. data/hard/.DS_Store +0 -0
  42. data/hard/seed_0/audit_summary.csv +2 -0
  43. data/hard/seed_0/config.yaml +30 -0
  44. data/hard/seed_0/matched_pairs.parquet +3 -0
  45. data/hard/seed_0/schema.json +84 -0
  46. data/hard/seed_0/transactions.parquet +3 -0
  47. data/hard/seed_1/audit_summary.csv +2 -0
  48. data/hard/seed_1/config.yaml +30 -0
  49. data/hard/seed_1/matched_pairs.parquet +3 -0
  50. data/hard/seed_1/schema.json +84 -0
DATASET_CARD.md ADDED
@@ -0,0 +1,458 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Temporal Twins Dataset Card
2
+
3
+ ## 1. Dataset Summary
4
+
5
+ Temporal Twins is a synthetic UPI-style transaction benchmark for temporal fraud detection. It is designed to evaluate whether a model can distinguish fraud from benign behavior using order-sensitive temporal structure rather than static aggregates such as total transaction count, account age, or prefix length.
6
+
7
+ The benchmark simulates users sending transactions over time and then assigns fraud labels through delayed temporal mechanisms. Its core design is a matched fraud/benign temporal-twin construction:
8
+
9
+ - each positive example is a fraud twin evaluated at a local event index `k`
10
+ - each negative example is a benign twin evaluated at the same local event index `k`
11
+ - both twins are matched on static and prefix-level summaries
12
+ - the benign twin contains the same unordered ingredients but violates the fraud-relevant temporal order
13
+
14
+ Temporal Twins exposes four benchmark modes:
15
+
16
+ - `oracle_calib`
17
+ - `easy`
18
+ - `medium`
19
+ - `hard`
20
+
21
+ The frozen paper-suite configuration used in this repository is:
22
+
23
+ - `num_users = 350`
24
+ - `simulation_days = 45`
25
+ - `seeds = [0, 1, 2, 3, 4]`
26
+ - `fast_mode = false`
27
+ - `n_checkpoints = 8`
28
+
29
+ ## 2. Dataset Motivation
30
+
31
+ Many fraud datasets can be solved by static shortcuts: longer histories, later evaluation times, higher transaction counts, or other aggregate correlates can make a benchmark look temporally rich while actually rewarding non-temporal models. Temporal Twins was built to remove those shortcuts and isolate order-sensitive temporal reasoning.
32
+
33
+ The benchmark therefore aims to answer a narrower research question:
34
+
35
+ - when static summaries are matched between positives and negatives, can a model still recover delayed fraud signals from temporal order alone?
36
+
37
+ It is intended for benchmarking temporal representation learning, causal order sensitivity, and delayed-label detection under controlled synthetic conditions.
38
+
39
+ ## 3. Dataset Composition
40
+
41
+ Temporal Twins is generated programmatically from synthetic user and transaction processes. There is no fixed real-world corpus. Each generated artifact is an event table in which each row is a synthetic transaction.
42
+
43
+ At a high level, each run contains:
44
+
45
+ - a synthetic user population
46
+ - a synthetic stream of UPI-style transactions
47
+ - risk-engine outputs such as transaction risk scores and failures
48
+ - benchmark-specific fraud and audit annotations
49
+ - matched fraud/benign evaluation pairs extracted from the event stream
50
+
51
+ The paper-scale suite in this repository contains 20 deterministic runs:
52
+
53
+ - `oracle_calib` with seeds `0..4`
54
+ - `easy` with seeds `0..4`
55
+ - `medium` with seeds `0..4`
56
+ - `hard` with seeds `0..4`
57
+
58
+ Mean matched evaluation-pair counts in the frozen paper suite are:
59
+
60
+ | Mode | Matched evaluation pairs (mean +- std) |
61
+ |---|---:|
62
+ | `oracle_calib` | `2606.6 +- 454.3` |
63
+ | `easy` | `2222.2 +- 128.4` |
64
+ | `medium` | `2356.6 +- 18.0` |
65
+ | `hard` | `2317.6 +- 22.0` |
66
+
67
+ Each paper-suite run is class-balanced at evaluation time:
68
+
69
+ - positives = negatives
70
+ - positive rate = `0.5000`
71
+
72
+ ## 4. Dataset Generation Process
73
+
74
+ The generation pipeline has four stages:
75
+
76
+ 1. Synthetic user generation
77
+ 2. Synthetic transaction generation
78
+ 3. Synthetic risk and retry generation
79
+ 4. Fraud-mechanism and matched-twin generation
80
+
81
+ More concretely:
82
+
83
+ 1. A synthetic user set is created with user-level behavioral parameters.
84
+ 2. A synthetic transaction stream is sampled with sender IDs, receiver IDs, timestamps, transaction amounts, and transaction types.
85
+ 3. A risk engine adds synthetic risk-related fields such as `risk_score`, `fail_prob`, `failed`, and retry-like events.
86
+ 4. The fraud engine applies benchmark-mode-specific temporal mechanisms and constructs matched temporal twins.
87
+
88
+ For the `temporal_twins` benchmark family, the generator then:
89
+
90
+ - constructs fraud twins and benign twins from matched carrier users and templates
91
+ - preserves matched static and prefix-level summaries
92
+ - injects delayed fraud labels into fraud twins
93
+ - forces benign twins to avoid the fraud-relevant temporal motif while retaining similar unordered ingredients
94
+
95
+ The benchmark is deterministic under fixed configuration, seed, and runtime settings.
96
+
97
+ ## 5. Fraud Mechanisms
98
+
99
+ Temporal Twins uses delayed, order-sensitive fraud mechanisms rather than directly labeling static outliers. Important mechanisms include:
100
+
101
+ - velocity-like activity acceleration
102
+ - retry-like behavior
103
+ - delayed receiver revisits
104
+ - burst-release-burst motifs
105
+ - adversarial timing perturbations
106
+ - delayed fraud assignment
107
+ - hidden latent fraud-state dynamics
108
+
109
+ These mechanisms are combined with difficulty-dependent noise and camouflage. In the standard `easy`, `medium`, and `hard` modes, the fraud signal is intentionally imperfect and partially obscured. In `oracle_calib`, the construction is designed to validate motif and evaluation alignment under matched-prefix conditions.
110
+
111
+ ## 6. Matched-Control Construction
112
+
113
+ The central benchmark control is the fraud/benign temporal twin.
114
+
115
+ For every fraud twin positive label at local event index `k`:
116
+
117
+ - the benign twin is evaluated at the same local event index `k`
118
+ - both examples use the same local prefix length
119
+ - both examples are truncated at prefix index `k`
120
+ - no future events are visible to the model
121
+
122
+ Within each matched pair, the protocol additionally matches:
123
+
124
+ - total transaction count
125
+ - local prefix length
126
+ - evaluation timestamp
127
+ - account age
128
+ - active age
129
+ - receiver histograms
130
+ - static aggregate summaries
131
+
132
+ In words:
133
+
134
+ - the fraud twin contains a temporally meaningful order pattern that triggers a delayed positive label
135
+ - the benign twin contains comparable ingredients and prefix statistics but violates the fraud-relevant temporal order
136
+
137
+ This design is meant to prevent performance from arising from:
138
+
139
+ - longer histories
140
+ - older accounts
141
+ - later prefix positions
142
+ - different transaction totals
143
+ - unmatched prefix ages
144
+ - benign negatives evaluated at arbitrary or easier positions
145
+
146
+ ## 7. Dataset Modes and Difficulty Ladder
147
+
148
+ Temporal Twins provides four modes.
149
+
150
+ ### `oracle_calib`
151
+
152
+ This is the calibration mode used to validate that the matched-prefix protocol is working as intended.
153
+
154
+ - Oracle metrics remain near-perfect.
155
+ - Static shortcut baselines remain at chance.
156
+ - Benign motif hit rate remains zero.
157
+ - This mode is primarily for protocol validation rather than realistic difficulty.
158
+
159
+ ### `easy`
160
+
161
+ - strong motif signal
162
+ - low noise
163
+ - shorter delay
164
+ - expected SeqGRU performance near `0.90-1.00`
165
+
166
+ ### `medium`
167
+
168
+ - moderate motif signal
169
+ - moderate noise
170
+ - longer delay
171
+ - expected SeqGRU performance near `0.80-0.90`
172
+
173
+ ### `hard`
174
+
175
+ - weaker motif signal
176
+ - longer delay
177
+ - adversarial perturbations and decoys
178
+ - expected SeqGRU performance near `0.70-0.85`
179
+
180
+ Naming convention:
181
+
182
+ - in `oracle_calib`, `AuditOracle` and `RawMotifOracle` are true oracle-style references
183
+ - in standard `easy`, `medium`, and `hard`, the corresponding scores are reported as `MotifProbe` and `RawMotifProbe` because realism and noise make them probes rather than perfect oracles
184
+
185
+ ## 8. Data Schema
186
+
187
+ The event table contains model-facing fields, supervision labels, and audit/oracle-only fields. The table below lists the most important columns used in this repository.
188
+
189
+ | Column name | Type | Description | Exposed to ordinary models? | Notes |
190
+ |---|---|---|---|---|
191
+ | `txn_id` | `int32` | Synthetic transaction identifier | Yes | Identifier only; not a benchmark target |
192
+ | `sender_id` | `int32` / `int64` | Synthetic sender account ID | Yes | Node identity available to temporal models |
193
+ | `receiver_id` | `int32` / `int64` | Synthetic receiver account ID | Yes | Used for graph and sequence structure |
194
+ | `timestamp` | `float32` | Synthetic event time in seconds from simulation start | Yes | Prefix truncation is based on timestamp and local index |
195
+ | `amount` | `float32` | Synthetic transaction amount | Yes | Not tied to real currency records |
196
+ | `txn_type` | `int8` | Synthetic transaction-type code | Yes | UPI-style categorical event attribute |
197
+ | `risk_score` | `float32` | Synthetic risk score from the risk engine | Yes | No real production risk model is used |
198
+ | `fail_prob` | `float32` | Synthetic failure probability | Yes | Risk-engine output |
199
+ | `failed` | `int8` | Binary failure indicator | Yes | Used as a normal model-facing field |
200
+ | `is_retry` | `int8` / derived | Retry-like event indicator | Yes | Available to ordinary models when present |
201
+ | `pair_freq` | `float32` / derived | Sender-receiver interaction-frequency feature | Yes | Derived from visible event history |
202
+ | `risk_noisy` | `float32` | Noisy synthetic risk feature | Yes | Benchmark feature, not an audit signal |
203
+ | `txn_count_10` | `float32` / derived | Recent-count feature over a short window | Yes | Derived from visible history |
204
+ | `amount_sum_10` | `float32` / derived | Recent amount-sum feature | Yes | Derived from visible history |
205
+ | `is_fraud` | `int8` | Binary fraud label | No | Supervision target only, not a model input |
206
+ | `twin_pair_id` | `int64` | Matched fraud/benign pair identifier | No | Audit/oracle-only; not exposed to learned baselines |
207
+ | `twin_role` | `string` | Twin role such as `fraud`, `benign`, or `background` | No | Audit/oracle-only |
208
+ | `twin_label` | `int8` | Pairwise matched label for audit utilities | No | Audit/oracle-only |
209
+ | `template_id` | `int64` | Source template identifier used during twin construction | No | Audit/oracle-only |
210
+ | `dynamic_fraud_state` | `float32` | Latent synthetic fraud-state variable | No | Hidden mechanism for analysis only |
211
+ | `motif_source` | `int8` | Indicator for motif-source events in a sequence | No | Audit/oracle-only |
212
+ | `motif_hit_count` | `int32` | Count of motif hits in the sequence | No | Audit/oracle-only |
213
+ | `trigger_event_idx` | `int32` | Local event index of the trigger event | No | Audit/oracle-only |
214
+ | `label_event_idx` | `int32` | Local event index at which the fraud label becomes active | No | Audit/oracle-only |
215
+ | `label_delay` | `int32` | Delay between trigger and labeled event index | No | Audit/oracle-only |
216
+ | `fraud_source` | `string` | Cause of fraud label, e.g. motif or fallback chain | No | Audit/oracle-only |
217
+ | `is_fallback_label` | `int8` | Indicator that a label came from fallback logic | No | Audit/oracle-only |
218
+ | `motif_chain_state` | `float32` | Internal motif-chain analysis field | No | Audit/oracle-only |
219
+ | `motif_strength` | `float32` | Internal motif-strength analysis field | No | Audit/oracle-only |
220
+
221
+ Not every baseline uses every model-facing column. The important guarantee is that learned baselines do not receive the audit/oracle-only fields listed above.
222
+
223
+ ## 9. Model-Facing vs Audit/Oracle-Only Columns
224
+
225
+ Ordinary learned baselines are restricted to model-facing transaction attributes and histories. In this repository, audit/oracle-only columns are explicitly stripped before learned baselines are trained or evaluated.
226
+
227
+ Ordinary models may use fields such as:
228
+
229
+ - `sender_id`
230
+ - `receiver_id`
231
+ - `timestamp`
232
+ - `amount`
233
+ - `risk_score`
234
+ - `fail_prob`
235
+ - `failed`
236
+ - `txn_type`
237
+ - other derived non-oracle features built from visible prefix history
238
+
239
+ Ordinary models must not use:
240
+
241
+ - `motif_hit_count`
242
+ - `motif_source`
243
+ - `trigger_event_idx`
244
+ - `label_event_idx`
245
+ - `label_delay`
246
+ - `fraud_source`
247
+ - `twin_role`
248
+ - `twin_label`
249
+ - `twin_pair_id`
250
+ - `template_id`
251
+ - `dynamic_fraud_state`
252
+ - other oracle-only diagnostics
253
+
254
+ This separation is necessary for the benchmark claim that performance should come from temporal reasoning rather than privileged audit information.
255
+
256
+ ## 10. Benchmark Tasks
257
+
258
+ Temporal Twins supports the following benchmark task:
259
+
260
+ - binary fraud detection on matched prefix examples
261
+
262
+ The standard evaluation protocol is:
263
+
264
+ - build matched fraud/benign examples
265
+ - truncate each sender history at the matched prefix index `k`
266
+ - train or score on the visible prefix only
267
+ - evaluate binary discrimination at the matched example level
268
+
269
+ Primary reported metrics include:
270
+
271
+ - ROC-AUC
272
+ - PR-AUC
273
+ - shuffled-order ROC-AUC
274
+ - shuffle delta = shuffled ROC-AUC minus clean ROC-AUC
275
+
276
+ The shuffled-order test is important: it measures how much performance depends on event order rather than unordered ingredients.
277
+
278
+ ## 11. Baselines and Reference Results
279
+
280
+ The frozen 5-seed paper suite uses:
281
+
282
+ - `num_users = 350`
283
+ - `simulation_days = 45`
284
+ - `seeds = [0, 1, 2, 3, 4]`
285
+ - `fast_mode = false`
286
+ - `n_checkpoints = 8`
287
+
288
+ Compact reference results:
289
+
290
+ | Mode | Primary reference | Secondary reference | XGBoost ROC-AUC | StaticGNN ROC-AUC | SeqGRU ROC-AUC | SeqGRU shuffled delta |
291
+ |---|---:|---:|---:|---:|---:|---:|
292
+ | `oracle_calib` | `AuditOracle 1.0000 +- 0.0000` | `RawMotifOracle 1.0000 +- 0.0000` | `0.5000 +- 0.0000` | `0.5222 +- 0.0235` | `1.0000 +- 0.0000` | `-0.5032 +- 0.0043` |
293
+ | `easy` | `MotifProbe 1.0000 +- 0.0000` | `RawMotifProbe 0.9983 +- 0.0011` | `0.5000 +- 0.0000` | `0.4946 +- 0.0128` | `1.0000 +- 0.0000` | `-0.5003 +- 0.0096` |
294
+ | `medium` | `MotifProbe 0.6374 +- 0.0069` | `RawMotifProbe 0.6482 +- 0.0058` | `0.5000 +- 0.0000` | `0.4922 +- 0.0203` | `0.8391 +- 0.0174` | `-0.3337 +- 0.0191` |
295
+ | `hard` | `MotifProbe 0.5790 +- 0.0045` | `RawMotifProbe 0.5910 +- 0.0105` | `0.5000 +- 0.0000` | `0.5026 +- 0.0198` | `0.6876 +- 0.0128` | `-0.1883 +- 0.0111` |
296
+
297
+ Static shortcut audit across all 20 paper-suite runs:
298
+
299
+ - `static_agg_auc = 0.5000 +- 0.0000`
300
+ - `total_txn_count AUC = 0.5000 +- 0.0000`
301
+ - `local_event_idx AUC = 0.5000 +- 0.0000`
302
+ - `prefix_txn_count AUC = 0.5000 +- 0.0000`
303
+ - `timestamp AUC = 0.5000 +- 0.0000`
304
+ - `account_age AUC = 0.5000 +- 0.0000`
305
+ - `active_age AUC = 0.5000 +- 0.0000`
306
+ - `benign_motif_hit_rate = 0.0000 +- 0.0000`
307
+
308
+ These results support the intended interpretation:
309
+
310
+ - static shortcuts are neutralized
311
+ - `oracle_calib` validates matched-prefix correctness
312
+ - `easy` is readily learnable by order-sensitive sequence models
313
+ - `medium` remains learnable but meaningfully harder
314
+ - `hard` remains above static baselines but is substantially more challenging
315
+
316
+ Full paper-suite artifacts, including temporal GNN results and per-seed CSVs, are stored under:
317
+
318
+ - `results/paper_suite_20260503_202810/`
319
+
320
+ ## 12. Intended Use
321
+
322
+ This dataset is intended for:
323
+
324
+ - research on temporal fraud detection
325
+ - benchmarking order-sensitive sequence and temporal-graph models
326
+ - evaluating whether performance survives matched static controls
327
+ - studying delayed labels and prefix-only evaluation
328
+ - comparing clean-order and shuffled-order performance
329
+
330
+ It is appropriate for methodology papers, controlled ablation studies, and robustness checks on temporal inductive bias.
331
+
332
+ ## 13. Out-of-Scope Use
333
+
334
+ Temporal Twins is out of scope for:
335
+
336
+ - direct training of production fraud systems
337
+ - making real financial, banking, or payment decisions
338
+ - approving or denying transactions for real users
339
+ - risk-scoring real individuals or organizations
340
+ - regulatory, legal, or operational decisions in production financial systems
341
+
342
+ The dataset must not be used to train production fraud systems directly or to make real financial decisions.
343
+
344
+ ## 14. Limitations
345
+
346
+ Important limitations include:
347
+
348
+ - the benchmark is fully synthetic and reflects designer assumptions
349
+ - user behavior, fraud behavior, and benign behavior are simplified relative to real financial ecosystems
350
+ - the only ground truth is the generator's own labeling logic
351
+ - real-world fraud often depends on richer institutional, device, merchant, and social context not present here
352
+ - difficulty levels are benchmark design choices, not calibrated measures of real operational difficulty
353
+ - temporal GNN underperformance on this benchmark should not be generalized to all real fraud settings
354
+
355
+ ## 15. Biases and Risks
356
+
357
+ As a synthetic benchmark, Temporal Twins inherits the modeling biases of its generator:
358
+
359
+ - it emphasizes order-sensitive motifs chosen by the benchmark designers
360
+ - it encodes a particular notion of delayed fraud and camouflage
361
+ - it may reward models that are well aligned to these synthetic mechanisms
362
+ - it may underrepresent other real fraud styles not captured by the generator
363
+
364
+ There is also a scientific risk:
365
+
366
+ - because the benchmark intentionally removes common static shortcuts, performance on Temporal Twins may differ from performance on operational datasets where those shortcuts exist, for better or worse
367
+
368
+ ## 16. Privacy and Sensitive Data
369
+
370
+ Temporal Twins contains no real financial or personal data.
371
+
372
+ Specifically:
373
+
374
+ - no real UPI data
375
+ - no real users
376
+ - no real bank accounts
377
+ - no real transactions
378
+ - no personal financial records
379
+ - no protected demographic attributes
380
+
381
+ All user IDs, receiver IDs, timestamps, amounts, and risk signals are synthetic artifacts produced by the generator.
382
+
383
+ ## 17. Ethical Considerations
384
+
385
+ Temporal Twins is safer to share than real financial logs because it does not contain real persons or institutions. However, ethical care is still needed.
386
+
387
+ Users of the dataset should not:
388
+
389
+ - present synthetic results as direct evidence of production readiness
390
+ - claim fairness or social validity that has not been tested on real populations
391
+ - use the dataset as justification for automated decisions about real people
392
+
393
+ The intended ethical use is research benchmarking, not operational deployment.
394
+
395
+ ## 18. Reproducibility
396
+
397
+ The repository includes deterministic generation and evaluation settings for the frozen paper suite.
398
+
399
+ Paper-suite configuration:
400
+
401
+ - `num_users = 350`
402
+ - `simulation_days = 45`
403
+ - `seeds = [0, 1, 2, 3, 4]`
404
+ - `fast_mode = false`
405
+ - `n_checkpoints = 8`
406
+
407
+ Reproducibility properties:
408
+
409
+ - stable deterministic seed derivation is used for benchmark modes and profiles
410
+ - Python, NumPy, and PyTorch seeds are fixed per run
411
+ - deterministic runtime flags are enabled where safe
412
+ - matched-prefix datasets are reproducible under fixed config and seed
413
+ - the final paper suite in this repository is stored as deterministic CSV artifacts
414
+
415
+ Reference artifacts:
416
+
417
+ - `results/paper_suite_20260503_202810/paper_suite_runs.csv`
418
+ - `results/paper_suite_20260503_202810/paper_suite_summary.csv`
419
+ - `results/paper_suite_20260503_202810/paper_suite_runtime.csv`
420
+ - `results/paper_suite_20260503_202810/paper_suite_failed_checks.csv`
421
+
422
+ ## 19. Hosting, License, and Citation
423
+
424
+ ### Hosting
425
+
426
+ The benchmark is currently generated from code in this repository rather than distributed as a fixed external archive.
427
+
428
+ Current status:
429
+
430
+ - dataset hosting location: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins)
431
+ - canonical pre-generated release archive: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins)
432
+ - Croissant metadata URL: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/metadata/temporal_twins_croissant.json](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/metadata/temporal_twins_croissant.json)
433
+ - Croissant metadata browser page: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/blob/main/metadata/temporal_twins_croissant.json](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/blob/main/metadata/temporal_twins_croissant.json)
434
+ - data files: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/data](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/data)
435
+ - results: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/results](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/results)
436
+ - configs: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/configs](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/configs)
437
+ - metadata directory: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/metadata](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/metadata)
438
+ - reference paper-suite results: `results/paper_suite_20260503_202810/`
439
+
440
+ ### License
441
+
442
+ - Dataset license: `CC BY 4.0` (`CC-BY-4.0`)
443
+ - Code license: `Apache License 2.0` (`Apache-2.0`)
444
+
445
+ ### Citation
446
+
447
+ `TODO` placeholder BibTeX:
448
+
449
+ ```bibtex
450
+ @dataset{temporal_twins_todo,
451
+ title = {Temporal Twins: A Synthetic UPI-Style Benchmark for Temporal Fraud Detection},
452
+ author = {TODO},
453
+ year = {TODO},
454
+ howpublished = {TODO},
455
+ note = {Synthetic matched-prefix temporal fraud benchmark},
456
+ url = {TODO}
457
+ }
458
+ ```
LICENSE ADDED
@@ -0,0 +1,203 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ SPDX-License-Identifier: Apache-2.0
2
+
3
+ Apache License
4
+ Version 2.0, January 2004
5
+ https://www.apache.org/licenses/LICENSE-2.0
6
+
7
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
8
+
9
+ 1. Definitions.
10
+
11
+ "License" shall mean the terms and conditions for use, reproduction,
12
+ and distribution as defined by Sections 1 through 9 of this document.
13
+
14
+ "Licensor" shall mean the copyright owner or entity authorized by
15
+ the copyright owner that is granting the License.
16
+
17
+ "Legal Entity" shall mean the union of the acting entity and all
18
+ other entities that control, are controlled by, or are under common
19
+ control with that entity. For the purposes of this definition,
20
+ "control" means (i) the power, direct or indirect, to cause the
21
+ direction or management of such entity, whether by contract or
22
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
23
+ outstanding shares, or (iii) beneficial ownership of such entity.
24
+
25
+ "You" (or "Your") shall mean an individual or Legal Entity
26
+ exercising permissions granted by this License.
27
+
28
+ "Source" form shall mean the preferred form for making modifications,
29
+ including but not limited to software source code, documentation
30
+ source, and configuration files.
31
+
32
+ "Object" form shall mean any form resulting from mechanical
33
+ transformation or translation of a Source form, including but
34
+ not limited to compiled object code, generated documentation,
35
+ and conversions to other media types.
36
+
37
+ "Work" shall mean the work of authorship, whether in Source or
38
+ Object form, made available under the License, as indicated by a
39
+ copyright notice that is included in or attached to the work
40
+ (an example is provided in the Appendix below).
41
+
42
+ "Derivative Works" shall mean any work, whether in Source or Object
43
+ form, that is based on (or derived from) the Work and for which the
44
+ editorial revisions, annotations, elaborations, or other modifications
45
+ represent, as a whole, an original work of authorship. For the purposes
46
+ of this License, Derivative Works shall not include works that remain
47
+ separable from, or merely link (or bind by name) to the interfaces of,
48
+ the Work and Derivative Works thereof.
49
+
50
+ "Contribution" shall mean any work of authorship, including
51
+ the original version of the Work and any modifications or additions
52
+ to that Work or Derivative Works thereof, that is intentionally
53
+ submitted to Licensor for inclusion in the Work by the copyright owner
54
+ or by an individual or Legal Entity authorized to submit on behalf of
55
+ the copyright owner. For the purposes of this definition, "submitted"
56
+ means any form of electronic, verbal, or written communication sent
57
+ to the Licensor or its representatives, including but not limited to
58
+ communication on electronic mailing lists, source code control systems,
59
+ and issue tracking systems that are managed by, or on behalf of, the
60
+ Licensor for the purpose of discussing and improving the Work, but
61
+ excluding communication that is conspicuously marked or otherwise
62
+ designated in writing by the copyright owner as "Not a Contribution."
63
+
64
+ "Contributor" shall mean Licensor and any individual or Legal Entity
65
+ on behalf of whom a Contribution has been received by Licensor and
66
+ subsequently incorporated within the Work.
67
+
68
+ 2. Grant of Copyright License. Subject to the terms and conditions of
69
+ this License, each Contributor hereby grants to You a perpetual,
70
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
71
+ copyright license to reproduce, prepare Derivative Works of,
72
+ publicly display, publicly perform, sublicense, and distribute the
73
+ Work and such Derivative Works in Source or Object form.
74
+
75
+ 3. Grant of Patent License. Subject to the terms and conditions of
76
+ this License, each Contributor hereby grants to You a perpetual,
77
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
78
+ (except as stated in this section) patent license to make, have made,
79
+ use, offer to sell, sell, import, and otherwise transfer the Work,
80
+ where such license applies only to those patent claims licensable
81
+ by such Contributor that are necessarily infringed by their
82
+ Contribution(s) alone or by combination of their Contribution(s)
83
+ with the Work to which such Contribution(s) was submitted. If You
84
+ institute patent litigation against any entity (including a
85
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
86
+ or a Contribution incorporated within the Work constitutes direct
87
+ or contributory patent infringement, then any patent licenses
88
+ granted to You under this License for that Work shall terminate
89
+ as of the date such litigation is filed.
90
+
91
+ 4. Redistribution. You may reproduce and distribute copies of the
92
+ Work or Derivative Works thereof in any medium, with or without
93
+ modifications, and in Source or Object form, provided that You
94
+ meet the following conditions:
95
+
96
+ (a) You must give any other recipients of the Work or
97
+ Derivative Works a copy of this License; and
98
+
99
+ (b) You must cause any modified files to carry prominent notices
100
+ stating that You changed the files; and
101
+
102
+ (c) You must retain, in the Source form of any Derivative Works
103
+ that You distribute, all copyright, patent, trademark, and
104
+ attribution notices from the Source form of the Work,
105
+ excluding those notices that do not pertain to any part of
106
+ the Derivative Works; and
107
+
108
+ (d) If the Work includes a "NOTICE" text file as part of its
109
+ distribution, then any Derivative Works that You distribute must
110
+ include a readable copy of the attribution notices contained
111
+ within such NOTICE file, excluding those notices that do not
112
+ pertain to any part of the Derivative Works, in at least one
113
+ of the following places: within a NOTICE text file distributed
114
+ as part of the Derivative Works; within the Source form or
115
+ documentation, if provided along with the Derivative Works; or,
116
+ within a display generated by the Derivative Works, if and
117
+ wherever such third-party notices normally appear. The contents
118
+ of the NOTICE file are for informational purposes only and
119
+ do not modify the License. You may add Your own attribution
120
+ notices within Derivative Works that You distribute, alongside
121
+ or as an addendum to the NOTICE text from the Work, provided
122
+ that such additional attribution notices cannot be construed
123
+ as modifying the License.
124
+
125
+ You may add Your own copyright statement to Your modifications and
126
+ may provide additional or different license terms and conditions
127
+ for use, reproduction, or distribution of Your modifications, or
128
+ for any such Derivative Works as a whole, provided Your use,
129
+ reproduction, and distribution of the Work otherwise complies with
130
+ the conditions stated in this License.
131
+
132
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
133
+ any Contribution intentionally submitted for inclusion in the Work
134
+ by You to the Licensor shall be under the terms and conditions of
135
+ this License, without any additional terms or conditions.
136
+ Notwithstanding the above, nothing herein shall supersede or modify
137
+ the terms of any separate license agreement you may have executed
138
+ with Licensor regarding such Contributions.
139
+
140
+ 6. Trademarks. This License does not grant permission to use the trade
141
+ names, trademarks, service marks, or product names of the Licensor,
142
+ except as required for reasonable and customary use in describing the
143
+ origin of the Work and reproducing the content of the NOTICE file.
144
+
145
+ 7. Disclaimer of Warranty. Unless required by applicable law or
146
+ agreed to in writing, Licensor provides the Work (and each
147
+ Contributor provides its Contributions) on an "AS IS" BASIS,
148
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
149
+ implied, including, without limitation, any warranties or conditions
150
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
151
+ PARTICULAR PURPOSE. You are solely responsible for determining the
152
+ appropriateness of using or redistributing the Work and assume any
153
+ risks associated with Your exercise of permissions under this License.
154
+
155
+ 8. Limitation of Liability. In no event and under no legal theory,
156
+ whether in tort (including negligence), contract, or otherwise,
157
+ unless required by applicable law (such as deliberate and grossly
158
+ negligent acts) or agreed to in writing, shall any Contributor be
159
+ liable to You for damages, including any direct, indirect, special,
160
+ incidental, or consequential damages of any character arising as a
161
+ result of this License or out of the use or inability to use the
162
+ Work (including but not limited to damages for loss of goodwill,
163
+ work stoppage, computer failure or malfunction, or any and all
164
+ other commercial damages or losses), even if such Contributor
165
+ has been advised of the possibility of such damages.
166
+
167
+ 9. Accepting Warranty or Additional Liability. While redistributing
168
+ the Work or Derivative Works thereof, You may choose to offer,
169
+ and charge a fee for, acceptance of support, warranty, indemnity,
170
+ or other liability obligations and/or rights consistent with this
171
+ License. However, in accepting such obligations, You may act only
172
+ on Your own behalf and on Your sole responsibility, not on behalf
173
+ of any other Contributor, and only if You agree to indemnify,
174
+ defend, and hold each Contributor harmless for any liability
175
+ incurred by, or claims asserted against, such Contributor by reason
176
+ of your accepting any such warranty or additional liability.
177
+
178
+ END OF TERMS AND CONDITIONS
179
+
180
+ APPENDIX: How to apply the Apache License to your work.
181
+
182
+ To apply the Apache License to your work, attach the following
183
+ boilerplate notice, with the fields enclosed by brackets "[]"
184
+ replaced with your own identifying information. (Do not include
185
+ the brackets.) The text should be enclosed in the appropriate
186
+ comment syntax for the file format. We also recommend that a
187
+ file or class name and description of purpose be included on the
188
+ same "printed page" as the copyright notice for easier
189
+ identification within third-party archives.
190
+
191
+ Copyright [yyyy] [name of copyright owner]
192
+
193
+ Licensed under the Apache License, Version 2.0 (the "License");
194
+ you may not use this file except in compliance with the License.
195
+ You may obtain a copy of the License at
196
+
197
+ https://www.apache.org/licenses/LICENSE-2.0
198
+
199
+ Unless required by applicable law or agreed to in writing, software
200
+ distributed under the License is distributed on an "AS IS" BASIS,
201
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
202
+ See the License for the specific language governing permissions and
203
+ limitations under the License.
LICENSE-DATA ADDED
@@ -0,0 +1,18 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ SPDX-License-Identifier: CC-BY-4.0
2
+
3
+ Temporal Twins dataset artifacts, generated synthetic data, metadata, dataset card,
4
+ and released benchmark files are licensed under the Creative Commons Attribution
5
+ 4.0 International license (CC BY 4.0).
6
+
7
+ Canonical license URL:
8
+ https://creativecommons.org/licenses/by/4.0/
9
+
10
+ This applies to released synthetic benchmark artifacts, including generated data
11
+ exports, metadata files, release bundle contents, and benchmark documentation that
12
+ describes the dataset.
13
+
14
+ Attribution requirement:
15
+ "If you use Temporal Twins, please cite the associated paper and dataset release."
16
+
17
+ Temporal Twins contains synthetic benchmark data only. It does not include real UPI
18
+ transactions, real users, real bank accounts, or personal financial records.
MANIFEST.sha256 ADDED
@@ -0,0 +1,126 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ b880f7dc777e417cd1a526b141b6e15866e89e62f8d4bfe627609aa909d40a65 .DS_Store
2
+ 6bebd0daf847a3885b372df91ca9bbb5548e7b84d162db9e1f3c6178af6a9465 DATASET_CARD.md
3
+ 15b001b0571fac1a30ea353be175df2724e7368d9c7ac9b433b9ac7afe2eb698 LICENSE
4
+ 6da8eaaf7b897c14b497468e485beae1b5c3d0514f1b1461e30133b890be996b LICENSE-DATA
5
+ 108e920e2e46828bed2239f8c204da0275454e18a2106904296813668a35d7bd README.md
6
+ 7d90d4c3ea8a567a4b264a46b28549d4d45a4350ee81493f1e7579f4e5aa5e8e README_REPO.md
7
+ 47125bd1c86a1c1d471fbfaef505a1a1af660374f537de6dac57aee622e3522f RELEASE_CHECKLIST.md
8
+ 7417423f35a909ebf8ed2e26b9124d00d677a0f10830b75db556c58dd3610b50 configs/default.yaml
9
+ c6433ed207c65b7be4ae74607f4badfff019f24ed2e5db0136c8c34f7b6d5d0c configs/paper_suite_reference.yaml
10
+ 5e1fc6d481fbdbe40c781302932ec8ab2813ec42ffd57c6cfd67550b7fb7cede configs/temporal_twins_calib.yaml
11
+ 13836465d6040605306f8d59635e408e6c96fa3ae96b862b82dadfa37f4428d0 croissant.json
12
+ 863ab315a86a9054f1c0f846d708f05452c53a40ae10fa654d5fbbe2065030b8 data/.DS_Store
13
+ b6be44f5edbcef72940bff042273a19811b890ca698509ff0a768d7ef1bcd572 data/README_GENERATION.md
14
+ 52a31f20444ef793ddcbf6705b8825846f66cae4d3150b41bb991191a07e577c data/_export_summary.csv
15
+ cf9ca3af6b4f9be5ed4b44409ff04175a1884ec616cecaedfdbe3a7d2675f009 data/easy/.DS_Store
16
+ c54c415f7ff4b08f08501e23a8ba013af03a0ccc407eadb3dc7b4478075f496f data/easy/seed_0/audit_summary.csv
17
+ e093124b5ba464725bbe1a673a74636a12a07b68fe202ad7bbc74fb4dbf15fbe data/easy/seed_0/config.yaml
18
+ d08d99bb61bfc154f2020f20458f4ba34084737ca12c191784d00e4c5f9a968b data/easy/seed_0/matched_pairs.parquet
19
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/easy/seed_0/schema.json
20
+ 8024a68ece6179d8d3af63ffd66eb8a444c79737760049dda816a03e4a25dbd7 data/easy/seed_0/transactions.parquet
21
+ c2b741e1c449f69bd356af02263b1bd0290af6ca5b9448fb5508e9e8d171b934 data/easy/seed_1/audit_summary.csv
22
+ 86d34ca36e4dfdeb353819227793c7610d0a9083c26de42e89eac73fe24a26c2 data/easy/seed_1/config.yaml
23
+ 7503b30b1b597adc9507fb10361d8ef0c0ca6e5add89c502843caba3a0aa9e74 data/easy/seed_1/matched_pairs.parquet
24
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/easy/seed_1/schema.json
25
+ 22f5283274a62e56a22458610709081d8b1bbf6e4793b6363c3ca9e518dc7f3e data/easy/seed_1/transactions.parquet
26
+ 480d45f3b0fded865d04f36877d5fcb500fc5ea878bbcf271ce39ea78e2bcf22 data/easy/seed_2/audit_summary.csv
27
+ d433c4686b343c8c921ebeae1cfd5d01d19fd3043ef26db9fa28736220d0a86a data/easy/seed_2/config.yaml
28
+ 583ee815999e1deade36980ead3d9f8cc93a23303e607b925728e153e2e250bd data/easy/seed_2/matched_pairs.parquet
29
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/easy/seed_2/schema.json
30
+ c61849d3d9679ef08fe1f595ced19145f2ecf60c15afc5c1d7998eb245b48f91 data/easy/seed_2/transactions.parquet
31
+ b27fa509489f00c3217547d482df01ca2f8d0d4a7b3cc25a27d75a8a2698eedf data/easy/seed_3/audit_summary.csv
32
+ 044f3f7eb0b4830050b0adb0ef10042dea1abaa3af2dfce17a3950d1878cb9c1 data/easy/seed_3/config.yaml
33
+ 6844de17f7fb80b858563672a15dfb38102e0dc1f370a9690b7c0f2ae5a66145 data/easy/seed_3/matched_pairs.parquet
34
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/easy/seed_3/schema.json
35
+ d4b28ff442286708fcc9fe214e53187c5a23d14f634b1387f4e2527c2e9ccd25 data/easy/seed_3/transactions.parquet
36
+ ae8988b49b91e054aa13ecda8d42364bebf990d5a6eea3d5f991a1b43c05d9b8 data/easy/seed_4/audit_summary.csv
37
+ dd509e61debe577b7cd982ae3aae43ba89ca21018ddda70b594f2cd7c3b8ef4b data/easy/seed_4/config.yaml
38
+ 086f1c0336b7a670b6279207751fc337cf7d0a30cad24d4ae1b98542e94a4884 data/easy/seed_4/matched_pairs.parquet
39
+ 529be5dcad73e43b738f4ce7a3b313ce96d533373b0901f361ae4e6e44006175 data/easy/seed_4/schema.json
40
+ f12641c7b1496ea8a84f9d26fccd9a65f9fdbd4678de0c821a472dfbb6ca9c58 data/easy/seed_4/transactions.parquet
41
+ f8c7ce63031d7bdba3e782908e0153733b45d4090dbf3f4036b6728cb98f662b data/hard/.DS_Store
42
+ b772c33cc025d18413ad3b67f294ce48dfb33004bd4efc453dca255a4c9317c8 data/hard/seed_0/audit_summary.csv
43
+ dd1cc092f279bc89d9036164d17c2084e026d67c476b267912a0dbf4b0d339c8 data/hard/seed_0/config.yaml
44
+ cd7ab1965e4b8961bf5a314170a7ed2a33211b5006ce2ac3fe27327d1cfdaf5d data/hard/seed_0/matched_pairs.parquet
45
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/hard/seed_0/schema.json
46
+ 5c6087b6fbbf6fe37ebf59e9e8c364c54ec1ebe6b6d7bd26add4ebcd819f68f0 data/hard/seed_0/transactions.parquet
47
+ 2ff501c60cf47dc00700320a563d9117b37c5ee498f18c6d61653e9d90099746 data/hard/seed_1/audit_summary.csv
48
+ 368b9e0e7c974ce73a0a04584c772aca03e3e806708f7787f81cab30826b4790 data/hard/seed_1/config.yaml
49
+ 69211d9c6d5c6d43b1b2e3ce897c152b2928bd7ff579032af63b398661888fa7 data/hard/seed_1/matched_pairs.parquet
50
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/hard/seed_1/schema.json
51
+ 1c5cc5cf873ee031971d46a6880b0d60a892c1833551cb134cbaabf1caf7f6b1 data/hard/seed_1/transactions.parquet
52
+ 8969c58b3baa5d72f83fd747dbd4ef2ff91917a6ca345583b02dd4b499548ae0 data/hard/seed_2/audit_summary.csv
53
+ 1c02d29959ba420021d93bf7ab93aadbece1b55548233bde16c59d6991716f6b data/hard/seed_2/config.yaml
54
+ 04321f86862b0e92dcd719c66f42b5e608fb0861af20e09db488353281aa4f64 data/hard/seed_2/matched_pairs.parquet
55
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/hard/seed_2/schema.json
56
+ 9c2e62c11b7355b23af24a64985077f19454a24dacaf1a35c52134d19c25af14 data/hard/seed_2/transactions.parquet
57
+ 89cba695bba08caccda38b5630a3a458080b418a2bba1b245a6fb6ca130c6cc2 data/hard/seed_3/audit_summary.csv
58
+ 4cee29d2afb4fc61282e8a07265c7a28275979a0789b88f7998470fb916156b3 data/hard/seed_3/config.yaml
59
+ a56fcbd44fe74eff14159955088e6db318749c734a6e092ff235cc29ad40cbe9 data/hard/seed_3/matched_pairs.parquet
60
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/hard/seed_3/schema.json
61
+ 32fc1ed13445ecb69e8bb41e0be50f6c6a5037c5425d0fc27b14276b3352bfe9 data/hard/seed_3/transactions.parquet
62
+ c80d9640280f26ec6366242ae46a42f014f1e0e869c478c4d58deaa878dd22da data/hard/seed_4/audit_summary.csv
63
+ 2e36e47498202fb382f1b6bb8f15ebe073e62a45c9675de882ce6b92cc489db5 data/hard/seed_4/config.yaml
64
+ 6ff1882fa3c868d6f4c9ca23ba924c0ff0edf7144df0edfff75d391e5ffa3b46 data/hard/seed_4/matched_pairs.parquet
65
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/hard/seed_4/schema.json
66
+ 6c1a0b613544ca33b65d944a03b6afc6cba5cc107922da2a3244e60e3c8de299 data/hard/seed_4/transactions.parquet
67
+ 4c6e951bda65a402ec21b4e364f0bcadb2aec4f90b0e6d7a6a88565f4f789f3f data/medium/seed_0/audit_summary.csv
68
+ bc4a60500e2181e48003d74837b5ef9176875099d667b00f00b811a90f7963aa data/medium/seed_0/config.yaml
69
+ 5733c743b7447498f8be2dd4dddb73ce3670c618293e91b80e5d34d8988ccda8 data/medium/seed_0/matched_pairs.parquet
70
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/medium/seed_0/schema.json
71
+ 46f6d1cf871f444d0690e0e56a64d7c13a2c333b2ba88a9f4179bdc48550c5f2 data/medium/seed_0/transactions.parquet
72
+ f90c79e62d5ee49fa6217a72e5374d5ad001c86bf6bdf4c66ae5ee65b014df22 data/medium/seed_1/audit_summary.csv
73
+ 2e7a495ef3366c9bcfbd5b6e3bccbcf4660244f1ed1679ec2f6d4324ac2fa9f3 data/medium/seed_1/config.yaml
74
+ 65a7925d2ff260e08b0aac41f160b10cedb1826f7ed4834dd247b54d23207808 data/medium/seed_1/matched_pairs.parquet
75
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/medium/seed_1/schema.json
76
+ 193179b04a363cdb76f5e7e65132099e3310d358610fb4a813099299fa54bf55 data/medium/seed_1/transactions.parquet
77
+ 501ec95ab1ce98a30fab0a1ea5d5d130a0a114aaef9e181bb511c7aec8e17260 data/medium/seed_2/audit_summary.csv
78
+ 9e9ea932f9916310f2376b531fecc4b147a7676bc240393ae1e9dea454843a8e data/medium/seed_2/config.yaml
79
+ 9a2b75dcb094d8b27ab26d99a9a68d4d55d392c1de2bd5024bc47620ab60bb2f data/medium/seed_2/matched_pairs.parquet
80
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/medium/seed_2/schema.json
81
+ 69d184da05bcfa79cd81223c4da12eef06daaad34f4845d31c229af2a6cb2915 data/medium/seed_2/transactions.parquet
82
+ dc2e1005f2eb25597826ff713713130ff07850dcb80d98a026244c1e9a71f664 data/medium/seed_3/audit_summary.csv
83
+ 2f7b88c2f9562b3638632586c0156edd1c1224fb775faf197e4d6d6d8fea0ce7 data/medium/seed_3/config.yaml
84
+ 022e2605912f071345c70337f6664f48f79d50cf572335e826979aca84f340c2 data/medium/seed_3/matched_pairs.parquet
85
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/medium/seed_3/schema.json
86
+ bb0917ec2b03f70808c5fddb1f11383aaf26427b2f2a8f20144ba51c8e46b648 data/medium/seed_3/transactions.parquet
87
+ b5ea04af6d9237151e7d0a4717b241e91596a84da5e1913d7a46dfe21bf0aeed data/medium/seed_4/audit_summary.csv
88
+ d6e4c69791f9d8b2fa1cdbb7fde5a18897812f914b6b90d96319b784740850e6 data/medium/seed_4/config.yaml
89
+ 42d684d91f8034dce83d20439016111cda53e5bb32582427ce6e89d04e39a876 data/medium/seed_4/matched_pairs.parquet
90
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/medium/seed_4/schema.json
91
+ f44c1f964fe6577c82cd46a291517238384c8797869981de166f0d2300fe27a6 data/medium/seed_4/transactions.parquet
92
+ 595d1aaba2fa576244f683e7b5d0b17bc26b1dc0319eac39e97b63ecfb76d834 data/oracle_calib/.DS_Store
93
+ 32d9b64974e2706a205dd7503545155a2cd2b2533988ca9f5f1c114e522ed478 data/oracle_calib/seed_0/audit_summary.csv
94
+ 9b88080413b4b51f689e0b63bf70f40b2b2a3f971144fd9de13a7ba6c7cc00fb data/oracle_calib/seed_0/config.yaml
95
+ 70356ba674fa717aa96c0527d35b36c5c617f56cc2f4c87f0a5667391679d5f7 data/oracle_calib/seed_0/matched_pairs.parquet
96
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/oracle_calib/seed_0/schema.json
97
+ ee194aefa2e121c0f4431eb1a071a2f858bac09171cc2165677f1b69a2c8a793 data/oracle_calib/seed_0/transactions.parquet
98
+ a5ad44bfb5d913734bd15352576a59777e18d12338a55ddfc81c233da44982f2 data/oracle_calib/seed_1/audit_summary.csv
99
+ a63c3ebd464e66c0a83e52c6234c350dbe37d5e5ddeb6d6db0bd2d372ef61151 data/oracle_calib/seed_1/config.yaml
100
+ 81088456e682c0176200deb20d67946f1bc5afc92f7991531b16c5c5a7b02ca9 data/oracle_calib/seed_1/matched_pairs.parquet
101
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/oracle_calib/seed_1/schema.json
102
+ 5db892f1c71f7f7b95df37e9201d1451f62e93948ba632bdcdae98a36803df8a data/oracle_calib/seed_1/transactions.parquet
103
+ 5de6202b1390b7869b506cdbc2947af80a4d1fb355acdbebf464b2cdd555ae79 data/oracle_calib/seed_2/audit_summary.csv
104
+ bb2febc8a3e748a8f8669d3f07fb62c4c9448b053429f89172c26839debdf0e6 data/oracle_calib/seed_2/config.yaml
105
+ 43e751fd2aad6043e931d5d246ad76b9c587bffeb91c5b913c5055bf593cb8e5 data/oracle_calib/seed_2/matched_pairs.parquet
106
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/oracle_calib/seed_2/schema.json
107
+ 3c50316eaa22cd9fcb2b3a37f0779cf31869f22edf35b05610f094b7f41b561f data/oracle_calib/seed_2/transactions.parquet
108
+ e8bbae36d063058af4becb3b4238b5bde091700f5896260573efb4f808582ad5 data/oracle_calib/seed_3/audit_summary.csv
109
+ f9ebdf69b21b9305551f8d0535096e8ce3d821aa5838199e76079efa91098039 data/oracle_calib/seed_3/config.yaml
110
+ 3e6e7e66b7a4516d9b3fb279c37aea20a6dd064d3e514f8db19c5e636e40a7b7 data/oracle_calib/seed_3/matched_pairs.parquet
111
+ b31dae63099e6684d521529f41acb740e2d6a826bb11204122ec08132bf48c4f data/oracle_calib/seed_3/schema.json
112
+ 9458d5f145e5a756c856979febee5dd6a4a049d5e59d801d63a5f0a208c65f9a data/oracle_calib/seed_3/transactions.parquet
113
+ 5d5cabc1a6d8e88997d9284484f44b14de30831cd24787866bea18176085cba4 data/oracle_calib/seed_4/audit_summary.csv
114
+ bef26c7c596322b6b0aab2199130d330d320bcda4ac1a478318ce4f6e25593b9 data/oracle_calib/seed_4/config.yaml
115
+ b53955a4f865446a5d1fef311fdde1209150f8a3116835841fd2352088b1592a data/oracle_calib/seed_4/matched_pairs.parquet
116
+ 529be5dcad73e43b738f4ce7a3b313ce96d533373b0901f361ae4e6e44006175 data/oracle_calib/seed_4/schema.json
117
+ 6974ebe8b33b719486fd44ccf39a08bf08dfda366ceaa34b65de2bc50ab7b01c data/oracle_calib/seed_4/transactions.parquet
118
+ 2debd2288dd3c5733c2c02fffbdd88983be77b978748574f1ab35ba8cb8a5612 metadata/CROISSANT_VALIDATION_NOTES.md
119
+ 13836465d6040605306f8d59635e408e6c96fa3ae96b862b82dadfa37f4428d0 metadata/temporal_twins_croissant.json
120
+ 6a79821117bdf30431ed79fa04da21df530553cd3ac22aeba2c58042afa79c0a results/PAPER_GATE_INTERPRETATION.md
121
+ 860940b8e6594ba0b44ecf8f68d09eb0fa9b32b5426de38be6ed094cdfe4b267 results/paper_suite_failed_checks.csv
122
+ 700e40d1ced465d61f681ad9c3f91c923f0770955967e27eac7ef9a5e99a0a6a results/paper_suite_meta.json
123
+ 1445666d207ab28d94678cdbf3625bf771700bdd1c444aa0cf01f41f6672055e results/paper_suite_runs.csv
124
+ 899415b8b34962cd1029b083a6f26282fe28402f03cd3877dd4da96d7840be74 results/paper_suite_runtime.csv
125
+ aabe56ba6dfcb585903b4df74c53fcbcdb82a0b48e75b1214232f2fa2daaa6e4 results/paper_suite_summary.csv
126
+ 839a448c5e8ab2e2c41647d2af607afd858a48f5ca8f213719bb0e480167c110 results/paper_suite_summary.md
README.md ADDED
@@ -0,0 +1,96 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Temporal Twins Release Bundle
2
+
3
+ This `release/` directory is a manual-hosting bundle for the Temporal Twins benchmark. It is intended to be uploaded later to a repository such as Hugging Face, Kaggle, Dataverse, OpenML, or another archival host. Hosting has **not** been performed by this preparation step.
4
+
5
+ ## What This Release Contains
6
+
7
+ - `DATASET_CARD.md`: NeurIPS-style dataset card for the benchmark
8
+ - `README_REPO.md`: copied repository README from the project root
9
+ - `LICENSE`: copied project license file
10
+ - `results/`: final deterministic paper-suite outputs
11
+ - `configs/`: benchmark configuration files and a paper-suite reference config
12
+ - `metadata/`: Croissant metadata and validation notes
13
+ - `data/`: empty per-mode/per-seed directory scaffold plus generation instructions
14
+ - `MANIFEST.sha256`: SHA256 manifest for all files in this bundle
15
+
16
+ ## How To Use The Data
17
+
18
+ 1. Read `DATASET_CARD.md` for benchmark scope, schema, intended use, and limitations.
19
+ 2. Read `metadata/temporal_twins_croissant.json` for machine-readable release metadata.
20
+ 3. Use `results/paper_suite_summary.csv` and `results/paper_suite_summary.md` for the paper-ready reference results.
21
+ 4. Populate `data/` with the generated per-seed transaction and matched-prefix files before public hosting.
22
+
23
+ ## How To Regenerate The Data
24
+
25
+ The repository does not currently store per-seed `transactions.parquet` and `matched_pairs.parquet` release exports. To generate them without changing benchmark logic, follow the instructions in `data/README_GENERATION.md`.
26
+
27
+ ## How To Reproduce Paper Results
28
+
29
+ The final deterministic paper suite used:
30
+
31
+ - benchmark groups: `oracle_calib`, `easy`, `medium`, `hard`
32
+ - benchmark modes:
33
+ - `oracle_calib` -> `temporal_twins_oracle_calib`
34
+ - `easy`, `medium`, `hard` -> `temporal_twins`
35
+ - seeds: `0 1 2 3 4`
36
+ - `num_users = 350`
37
+ - `simulation_days = 45`
38
+ - `fast_mode = false`
39
+ - `n_checkpoints = 8`
40
+ - device: `cpu`
41
+
42
+ Reference artifacts:
43
+
44
+ - `results/paper_suite_runs.csv`
45
+ - `results/paper_suite_summary.csv`
46
+ - `results/paper_suite_runtime.csv`
47
+ - `results/paper_suite_failed_checks.csv`
48
+ - `results/paper_suite_summary.md`
49
+ - `results/paper_suite_meta.json`
50
+
51
+ ## File Structure
52
+
53
+ ```text
54
+ release/
55
+ ├── README.md
56
+ ├── README_REPO.md
57
+ ├── DATASET_CARD.md
58
+ ├── LICENSE
59
+ ├── MANIFEST.sha256
60
+ ├── RELEASE_CHECKLIST.md
61
+ ├── data/
62
+ │ ├── README_GENERATION.md
63
+ │ ├── oracle_calib/seed_{0..4}/
64
+ │ ├── easy/seed_{0..4}/
65
+ │ ├── medium/seed_{0..4}/
66
+ │ └── hard/seed_{0..4}/
67
+ ├── configs/
68
+ ├── results/
69
+ └── metadata/
70
+ ```
71
+
72
+ ## Hosted URLs
73
+
74
+ - Dataset URL: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins)
75
+ - Croissant metadata URL: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/metadata/temporal_twins_croissant.json](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/metadata/temporal_twins_croissant.json)
76
+ - Croissant metadata browser page: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/blob/main/metadata/temporal_twins_croissant.json](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/blob/main/metadata/temporal_twins_croissant.json)
77
+ - Data URL: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/data](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/data)
78
+ - Results URL: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/results](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/results)
79
+ - Configs URL: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/configs](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/configs)
80
+ - Metadata URL: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/metadata](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/metadata)
81
+ - Full release archive: [https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins)
82
+ - Code repository: `TODO_CODE_REPOSITORY_URL`
83
+ - Paper or preprint: `TODO_PAPER_URL`
84
+
85
+ ## Licenses
86
+
87
+ - Code: Apache License 2.0 (`Apache-2.0`)
88
+ - Dataset and generated benchmark artifacts: Creative Commons Attribution 4.0 International (`CC-BY-4.0`)
89
+ - Code SPDX-License-Identifier: `Apache-2.0`
90
+ - Dataset SPDX-License-Identifier: `CC-BY-4.0`
91
+ - No real UPI data or personal financial records are included in this release bundle.
92
+
93
+ ## Notes
94
+
95
+ - This bundle contains no real UPI data, no real users, no real bank accounts, and no personal financial records.
96
+ - The benchmark code, generator logic, labels, matched-prefix protocol, and model logic were not modified while preparing this release directory.
README_REPO.md ADDED
@@ -0,0 +1,437 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Temporal Twins: A Matched-Control Benchmark for Temporal Fraud Detection
2
+
3
+ Temporal Twins is a synthetic UPI-style temporal transaction benchmark where fraud and benign trajectories are statically matched but differ in delayed event-order structure. The benchmark is designed to test whether models can exploit temporal ordering under matched-prefix controls rather than relying on static transaction summaries or prefix-length shortcuts.
4
+
5
+ ## Installation
6
+
7
+ Recommended Python version: `3.11+` (`3.13` also works in the checked environment).
8
+
9
+ ### pip
10
+
11
+ ```bash
12
+ pip install -r requirements.txt
13
+ ```
14
+
15
+ ### conda
16
+
17
+ If [environment.yml](environment.yml) is present:
18
+
19
+ ```bash
20
+ conda env create -f environment.yml
21
+ conda activate temporal-twins
22
+ ```
23
+
24
+ ## Repository Structure
25
+
26
+ - `src/`: synthetic user, transaction, risk, fraud, graph, and core config code
27
+ - `models/`: learned baselines and probe/oracle wrappers, including SeqGRU and temporal GNNs
28
+ - `experiments/`: benchmark runner and matched-prefix evaluation code
29
+ - `config/`: checked-in YAML configs used as base configs for experiments
30
+ - `results/`: frozen experiment artifacts, including the final deterministic paper suite
31
+ - `metadata/`: Croissant metadata and release-side validation notes
32
+ - `release/`: manual-hosting bundle prepared for later upload
33
+
34
+ ## Quick Smoke Test
35
+
36
+ The public CLI supports a fast audit-mode smoke test:
37
+
38
+ ```bash
39
+ PYTHONPATH=. python3 experiments/run_all.py \
40
+ --fast \
41
+ --seed 0 \
42
+ --benchmark-mode temporal_twins_oracle_calib \
43
+ --experiments audit \
44
+ --device cpu
45
+ ```
46
+
47
+ ## Exact Paper-Style Group Runner
48
+
49
+ The checked-in CLI does **not** expose `--difficulty`, `--num-users`, or `--simulation-days` flags. The exact grouped reproductions below therefore use the existing helper functions in [experiments/run_all.py](experiments/run_all.py) through an inline Python wrapper.
50
+
51
+ Define this shell helper once in your session:
52
+
53
+ ```bash
54
+ run_group() {
55
+ local group="$1"
56
+ local seed="$2"
57
+ local out_json="$3"
58
+
59
+ PYTHONPATH=. python3 - "$group" "$seed" "$out_json" <<'PY'
60
+ import json
61
+ import math
62
+ import sys
63
+ import time
64
+ from pathlib import Path
65
+
66
+ from src.core.config_loader import load_config
67
+ from experiments.run_all import (
68
+ build_gate_pool_from_frames,
69
+ gate_volume_is_sufficient,
70
+ generate_single_difficulty,
71
+ offset_gate_namespace,
72
+ prepare_gate_subset,
73
+ run_motif_validity_check,
74
+ set_global_determinism,
75
+ )
76
+
77
+
78
+ def normalize(value):
79
+ if isinstance(value, dict):
80
+ return {k: normalize(v) for k, v in value.items()}
81
+ if isinstance(value, (list, tuple)):
82
+ return [normalize(v) for v in value]
83
+ if hasattr(value, "item"):
84
+ try:
85
+ value = value.item()
86
+ except Exception:
87
+ pass
88
+ if isinstance(value, float) and not math.isfinite(value):
89
+ return None
90
+ return value
91
+
92
+
93
+ group = sys.argv[1]
94
+ seed = int(sys.argv[2])
95
+ out_json = Path(sys.argv[3])
96
+
97
+ if group == "oracle_calib":
98
+ benchmark_mode = "temporal_twins_oracle_calib"
99
+ difficulty = "easy"
100
+ hard_abort = True
101
+ force_temporal_models = True
102
+ else:
103
+ benchmark_mode = "temporal_twins"
104
+ difficulty = group
105
+ hard_abort = False
106
+ force_temporal_models = True
107
+
108
+ cfg = load_config("config/default.yaml")
109
+ cfg = cfg.model_copy(
110
+ update={
111
+ "num_users": 350,
112
+ "simulation_days": 45,
113
+ "benchmark_mode": benchmark_mode,
114
+ "random_seed": seed,
115
+ }
116
+ )
117
+
118
+ set_global_determinism(seed)
119
+ pool = generate_single_difficulty(
120
+ cfg,
121
+ difficulty=difficulty,
122
+ seed=seed,
123
+ benchmark_mode=benchmark_mode,
124
+ )
125
+ gate = prepare_gate_subset(pool, seed=seed, fast_mode=False)
126
+ pack_count = 1
127
+
128
+ while (not gate_volume_is_sufficient(gate["volume"], False)) and pack_count <= 6:
129
+ extra_seed = seed + pack_count * 10007
130
+ extra_pack = generate_single_difficulty(
131
+ cfg,
132
+ difficulty=difficulty,
133
+ seed=extra_seed,
134
+ benchmark_mode=benchmark_mode,
135
+ )
136
+ extra_pack = offset_gate_namespace(extra_pack, pack_count)
137
+ pool = build_gate_pool_from_frames([pool, extra_pack])
138
+ gate = prepare_gate_subset(pool, seed=seed, fast_mode=False)
139
+ pack_count += 1
140
+
141
+ gate["source_pool_events"] = int(len(pool))
142
+ gate["source_pool_pairs"] = int(pool.loc[pool["twin_pair_id"] >= 0, "twin_pair_id"].nunique()) if "twin_pair_id" in pool.columns else 0
143
+ gate["source_pool_packs"] = int(pack_count)
144
+
145
+ start = time.time()
146
+ gate_pass, report = run_motif_validity_check(
147
+ df=pool,
148
+ config=cfg,
149
+ seed=seed,
150
+ device="cpu",
151
+ num_epochs=3,
152
+ node_epochs=150,
153
+ n_checkpoints=8,
154
+ hard_abort=hard_abort,
155
+ benchmark_mode=benchmark_mode,
156
+ fast_mode=False,
157
+ force_temporal_models=force_temporal_models,
158
+ prebuilt_gate=gate,
159
+ )
160
+ elapsed = time.time() - start
161
+
162
+ result = {
163
+ "benchmark_group": group,
164
+ "benchmark_mode": benchmark_mode,
165
+ "seed": seed,
166
+ "primary_metric_label": report["audit_metric_label"],
167
+ "secondary_metric_label": report["raw_metric_label"],
168
+ "gate_pass": bool(gate_pass),
169
+ "run_wall_time_sec": float(elapsed),
170
+ **report,
171
+ }
172
+
173
+ out_json.parent.mkdir(parents=True, exist_ok=True)
174
+ out_json.write_text(json.dumps(normalize(result), indent=2) + "\n")
175
+ print(f"Wrote {out_json}")
176
+ PY
177
+ }
178
+ ```
179
+
180
+ ## Reproduce Oracle Calibration
181
+
182
+ Non-fast, reliable-volume `temporal_twins_oracle_calib`, seed `0`, `num_users=350`, `simulation_days=45`:
183
+
184
+ ```bash
185
+ run_group oracle_calib 0 results/paper_suite_repro/jobs/oracle_calib_0.json
186
+ ```
187
+
188
+ ## Reproduce Easy / Medium / Hard
189
+
190
+ Each command below reproduces the matched-prefix grouped benchmark for seed `0` with the paper-scale non-fast settings (`num_users=350`, `simulation_days=45`, `n_checkpoints=8`, deterministic CPU runtime):
191
+
192
+ ```bash
193
+ run_group easy 0 results/paper_suite_repro/jobs/easy_0.json
194
+ run_group medium 0 results/paper_suite_repro/jobs/medium_0.json
195
+ run_group hard 0 results/paper_suite_repro/jobs/hard_0.json
196
+ ```
197
+
198
+ ## Reproduce Full Paper Suite
199
+
200
+ There is no single checked-in `paper_suite` driver script. The exact grouped reproduction can be run as a shell loop over benchmark groups and seeds, followed by a small aggregation step that writes the artifact files:
201
+
202
+ ### 1. Generate per-run JSON files
203
+
204
+ ```bash
205
+ mkdir -p results/paper_suite_repro/jobs
206
+
207
+ for group in oracle_calib easy medium hard; do
208
+ for seed in 0 1 2 3 4; do
209
+ run_group "$group" "$seed" "results/paper_suite_repro/jobs/${group}_${seed}.json"
210
+ done
211
+ done
212
+ ```
213
+
214
+ ### 2. Aggregate into paper-suite CSV and Markdown files
215
+
216
+ ```bash
217
+ PYTHONPATH=. python3 - <<'PY'
218
+ import json
219
+ from pathlib import Path
220
+
221
+ import pandas as pd
222
+
223
+
224
+ def summarize_mean_std(df, group_col):
225
+ numeric_cols = [c for c in df.columns if c != group_col and pd.api.types.is_numeric_dtype(df[c])]
226
+ grouped = df.groupby(group_col, dropna=False)[numeric_cols].agg(["mean", "std"]).reset_index()
227
+ grouped.columns = [
228
+ group_col if col == group_col else f"{col}_{stat}"
229
+ for col, stat in grouped.columns
230
+ ]
231
+ return grouped
232
+
233
+
234
+ def volume_failures(row):
235
+ fails = []
236
+ if row["matched_eval_pairs"] < 2000:
237
+ fails.append(f"matched_eval_pairs={row['matched_eval_pairs']} (<2000)")
238
+ if row["positives"] < 500:
239
+ fails.append(f"positives={row['positives']} (<500)")
240
+ if row["negatives"] < 500:
241
+ fails.append(f"negatives={row['negatives']} (<500)")
242
+ if row["unique_fraud_users"] < 50:
243
+ fails.append(f"unique_fraud_users={row['unique_fraud_users']} (<50)")
244
+ if row["unique_benign_users"] < 50:
245
+ fails.append(f"unique_benign_users={row['unique_benign_users']} (<50)")
246
+ if not (0.35 <= row["positive_rate"] <= 0.65):
247
+ fails.append(f"positive_rate={row['positive_rate']:.4f} (outside [0.35,0.65])")
248
+ return " | ".join(fails)
249
+
250
+
251
+ def hard_gate_failures(row):
252
+ checks = [
253
+ (row["primary_metric_label"], row["audit_roc_auc"], ">=", 0.99),
254
+ (f"{row['primary_metric_label']} pair-sep", row["audit_pair_sep"], ">=", 0.99),
255
+ (row["secondary_metric_label"], row["raw_roc_auc"], ">=", 0.95),
256
+ (f"{row['secondary_metric_label']} pair-sep", row["raw_pair_sep"], ">=", 0.90),
257
+ ("static_agg_auc", row["static_agg_auc"], "<=", 0.60),
258
+ ("XGBoost ROC-AUC", row["xgb_roc_auc"], "<=", 0.65),
259
+ ("StaticGNN ROC-AUC", row["static_gnn_roc"], "<=", 0.70),
260
+ ("SeqGRU ROC-AUC", row["seqgru_roc_auc"], ">=", 0.80),
261
+ ("SeqGRU shuffle delta", row["seqgru_shuffle_delta"], "<=", -0.10),
262
+ ]
263
+ fails = []
264
+ for label, value, op, threshold in checks:
265
+ ok = value >= threshold if op == ">=" else value <= threshold
266
+ if not ok:
267
+ fails.append(f"{label}: {value:.4f} ({op}{threshold})")
268
+ return " | ".join(fails)
269
+
270
+
271
+ def advisory_failures(row):
272
+ checks = [
273
+ ("TGN ROC-AUC", row["tgn_roc_auc"], ">=", 0.75),
274
+ ("TGN shuffle delta", row["tgn_shuffle_delta"], "<=", -0.10),
275
+ ("TGAT ROC-AUC", row["tgat_roc_auc"], ">=", 0.75),
276
+ ("TGAT shuffle delta", row["tgat_shuffle_delta"], "<=", -0.10),
277
+ ("DyRep ROC-AUC", row["dyrep_roc_auc"], ">=", 0.75),
278
+ ("DyRep shuffle delta", row["dyrep_shuffle_delta"], "<=", -0.10),
279
+ ("JODIE ROC-AUC", row["jodie_roc_auc"], ">=", 0.75),
280
+ ("JODIE shuffle delta", row["jodie_shuffle_delta"], "<=", -0.10),
281
+ ]
282
+ fails = []
283
+ for label, value, op, threshold in checks:
284
+ ok = value >= threshold if op == ">=" else value <= threshold
285
+ if not ok:
286
+ fails.append(f"{label}: {value:.4f} ({op}{threshold})")
287
+ return " | ".join(fails)
288
+
289
+
290
+ jobs_dir = Path("results/paper_suite_repro/jobs")
291
+ out_dir = jobs_dir.parent
292
+ rows = [json.loads(path.read_text()) for path in sorted(jobs_dir.glob("*.json"))]
293
+ df = pd.DataFrame(rows).sort_values(["benchmark_group", "seed"]).reset_index(drop=True)
294
+
295
+ runs_path = out_dir / "paper_suite_runs.csv"
296
+ summary_path = out_dir / "paper_suite_summary.csv"
297
+ runtime_path = out_dir / "paper_suite_runtime.csv"
298
+ failed_path = out_dir / "paper_suite_failed_checks.csv"
299
+ summary_md_path = out_dir / "paper_suite_summary.md"
300
+ meta_path = out_dir / "paper_suite_meta.json"
301
+
302
+ df.to_csv(runs_path, index=False)
303
+
304
+ summary = summarize_mean_std(df, "benchmark_group")
305
+ summary.to_csv(summary_path, index=False)
306
+
307
+ runtime_cols = [
308
+ "benchmark_group",
309
+ "seed",
310
+ "run_wall_time_sec",
311
+ "static_gnn_eval_time_sec",
312
+ "static_gnn_unique_prefix_cutoffs",
313
+ "static_gnn_graph_builds",
314
+ "static_gnn_cache_hit_rate",
315
+ ]
316
+ df[runtime_cols].to_csv(runtime_path, index=False)
317
+
318
+ failed = df[["benchmark_group", "seed", "gate_pass"]].copy()
319
+ failed["volume_failures"] = df.apply(volume_failures, axis=1)
320
+ failed["hard_gate_failures"] = df.apply(hard_gate_failures, axis=1)
321
+ failed["advisory_failures"] = df.apply(advisory_failures, axis=1)
322
+ failed.to_csv(failed_path, index=False)
323
+
324
+ meta = {
325
+ "device": "cpu",
326
+ "num_users": 350,
327
+ "simulation_days": 45,
328
+ "num_epochs": 3,
329
+ "node_epochs": 150,
330
+ "n_checkpoints": 8,
331
+ "fast_mode": False,
332
+ "seeds": [0, 1, 2, 3, 4],
333
+ }
334
+ meta_path.write_text(json.dumps(meta, indent=2) + "\n")
335
+
336
+ headline = summary[
337
+ [
338
+ "benchmark_group",
339
+ "xgb_roc_auc_mean",
340
+ "static_gnn_roc_mean",
341
+ "seqgru_roc_auc_mean",
342
+ "seqgru_shuffle_delta_mean",
343
+ ]
344
+ ].copy()
345
+
346
+ lines = [
347
+ "# Paper Suite Summary",
348
+ "",
349
+ "| benchmark_group | xgb_roc_auc_mean | static_gnn_roc_mean | seqgru_roc_auc_mean | seqgru_shuffle_delta_mean |",
350
+ "|---|---:|---:|---:|---:|",
351
+ ]
352
+ for row in headline.itertuples(index=False):
353
+ lines.append(
354
+ f"| {row.benchmark_group} | {row.xgb_roc_auc_mean:.4f} | {row.static_gnn_roc_mean:.4f} | {row.seqgru_roc_auc_mean:.4f} | {row.seqgru_shuffle_delta_mean:.4f} |"
355
+ )
356
+ summary_md_path.write_text("\n".join(lines) + "\n")
357
+
358
+ print(f"Wrote {runs_path}")
359
+ print(f"Wrote {summary_path}")
360
+ print(f"Wrote {runtime_path}")
361
+ print(f"Wrote {failed_path}")
362
+ print(f"Wrote {summary_md_path}")
363
+ print(f"Wrote {meta_path}")
364
+ PY
365
+ ```
366
+
367
+ This aggregation step writes:
368
+
369
+ - `results/paper_suite_repro/paper_suite_runs.csv`
370
+ - `results/paper_suite_repro/paper_suite_summary.csv`
371
+ - `results/paper_suite_repro/paper_suite_runtime.csv`
372
+ - `results/paper_suite_repro/paper_suite_failed_checks.csv`
373
+ - `results/paper_suite_repro/paper_suite_summary.md`
374
+ - `results/paper_suite_repro/paper_suite_meta.json`
375
+
376
+ The frozen reference artifacts checked into this repository live in [results/paper_suite_20260503_202810](results/paper_suite_20260503_202810).
377
+
378
+ ## Expected Headline Results
379
+
380
+ | benchmark_group | XGBoost ROC-AUC | StaticGNN ROC-AUC | SeqGRU ROC-AUC | SeqGRU shuffle delta |
381
+ |---|---:|---:|---:|---:|
382
+ | `oracle_calib` | 0.5000 | 0.5222 | 1.0000 | -0.5032 |
383
+ | `easy` | 0.5000 | 0.4946 | 1.0000 | -0.5003 |
384
+ | `medium` | 0.5000 | 0.4922 | 0.8391 | -0.3337 |
385
+ | `hard` | 0.5000 | 0.5026 | 0.6876 | -0.1883 |
386
+
387
+ ## Determinism
388
+
389
+ - Deterministic CPU runtime is enabled in [experiments/run_all.py](experiments/run_all.py).
390
+ - The same seed should produce identical matched-prefix data and identical metrics under the same deterministic environment.
391
+ - Deterministic settings intentionally trade speed for repeatability and will slow larger runs.
392
+
393
+ For more detail, see [docs/DETERMINISM.md](docs/DETERMINISM.md).
394
+
395
+ ## Runtime Note
396
+
397
+ Mean wall-clock runtime per benchmark group in the final deterministic paper suite:
398
+
399
+ - `oracle_calib`: `1136.6s`
400
+ - `easy`: `1345.9s`
401
+ - `medium`: `2181.9s`
402
+ - `hard`: `2613.7s`
403
+ - cumulative summed runtime across all 20 runs: about `10.11` hours
404
+
405
+ ## Data and Metadata
406
+
407
+ - Dataset card: [DATASET_CARD.md](DATASET_CARD.md)
408
+ - Croissant metadata: [metadata/temporal_twins_croissant.json](metadata/temporal_twins_croissant.json)
409
+ - Manual-hosting release bundle: [release/](release/)
410
+
411
+ Hosted URLs:
412
+
413
+ - Dataset URL: [temporal-twins on Hugging Face](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins)
414
+ - Croissant metadata URL: [raw JSON-LD](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/metadata/temporal_twins_croissant.json)
415
+ - Croissant metadata browser page: [blob view](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/blob/main/metadata/temporal_twins_croissant.json)
416
+ - Data URL: [data tree](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/data)
417
+ - Results URL: [results tree](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/results)
418
+ - Configs URL: [configs tree](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/configs)
419
+ - Metadata URL: [metadata tree](https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/metadata)
420
+
421
+ ## License
422
+
423
+ - Code: Apache License 2.0 (`Apache-2.0`)
424
+ - Dataset and generated benchmark artifacts: Creative Commons Attribution 4.0 International (`CC-BY-4.0`)
425
+ - Code SPDX-License-Identifier: `Apache-2.0`
426
+ - Dataset SPDX-License-Identifier: `CC-BY-4.0`
427
+ - No real UPI data or personal financial records are included.
428
+
429
+ ## Citation
430
+
431
+ `TODO_REVEAL_AFTER_REVIEW`
432
+
433
+ ## Warning
434
+
435
+ - Synthetic data only
436
+ - No real UPI transactions
437
+ - Not for production fraud deployment
RELEASE_CHECKLIST.md ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Release Checklist
2
+
3
+ - Dataset card included.
4
+ - Croissant metadata included.
5
+ - Croissant dataset/data/results/configs/metadata URLs replaced after hosting.
6
+ - Code license selected: `Apache-2.0`.
7
+ - Dataset license selected: `CC-BY-4.0`.
8
+ - Code URL added.
9
+ - Dataset URL added.
10
+ - Manifest generated.
11
+ - No real user data.
12
+ - No secrets.
13
+ - Anonymized paths.
14
+
15
+ Hosted URLs:
16
+
17
+ - Dataset URL: `https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins`
18
+ - Croissant metadata URL: `https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/metadata/temporal_twins_croissant.json`
19
+ - Data URL: `https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/data`
20
+ - Results URL: `https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/results`
21
+ - Configs URL: `https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/configs`
22
+ - Metadata URL: `https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/metadata`
configs/default.yaml ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_users: 1000
2
+ simulation_days: 365
3
+ fraud_ratio: 0.05
4
+ benchmark_mode: temporal_twins
5
+
6
+ user_params:
7
+ lambda_mean: 5.0
8
+ lambda_std: 1.0
9
+ mu_mean: 7.5
10
+ mu_std: 1.0
11
+ sigma_mean: 0.5
12
+ sigma_std: 0.2
13
+
14
+ upi_limits:
15
+ max_txn_amount: 100000
16
+ daily_limit: 100000
17
+
18
+ risk_model:
19
+ weights:
20
+ amount_ratio: 1.0
21
+ daily_ratio: 0.8
22
+ velocity: 1.2
23
+ time_anomaly: 0.6
24
+ graph_anomaly: 1.0
25
+ retry: 0.8
26
+ kyc: 0.5
27
+ user_risk: 0.8
28
+
29
+ random_seed: 42
configs/paper_suite_reference.yaml ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ paper_suite:
2
+ benchmark_groups:
3
+ - oracle_calib
4
+ - easy
5
+ - medium
6
+ - hard
7
+ benchmark_modes:
8
+ oracle_calib: temporal_twins_oracle_calib
9
+ easy: temporal_twins
10
+ medium: temporal_twins
11
+ hard: temporal_twins
12
+ seeds:
13
+ - 0
14
+ - 1
15
+ - 2
16
+ - 3
17
+ - 4
18
+ num_users: 350
19
+ simulation_days: 45
20
+ fast_mode: false
21
+ n_checkpoints: 8
22
+ device: cpu
23
+ num_epochs: 3
24
+ node_epochs: 150
25
+ source_results_dir: results/paper_suite_20260503_202810
configs/temporal_twins_calib.yaml ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_users: 120
2
+ simulation_days: 30
3
+ fraud_ratio: 0.05
4
+ benchmark_mode: temporal_twins
5
+
6
+ user_params:
7
+ lambda_mean: 5.0
8
+ lambda_std: 1.0
9
+ mu_mean: 7.5
10
+ mu_std: 1.0
11
+ sigma_mean: 0.5
12
+ sigma_std: 0.2
13
+
14
+ upi_limits:
15
+ max_txn_amount: 100000
16
+ daily_limit: 100000
17
+
18
+ risk_model:
19
+ weights:
20
+ amount_ratio: 1.0
21
+ daily_ratio: 0.8
22
+ velocity: 1.2
23
+ time_anomaly: 0.6
24
+ graph_anomaly: 1.0
25
+ retry: 0.8
26
+ kyc: 0.5
27
+ user_risk: 0.8
28
+
29
+ random_seed: 42
croissant.json ADDED
@@ -0,0 +1,796 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "@context": {
3
+ "@vocab": "https://schema.org/",
4
+ "sc": "https://schema.org/",
5
+ "cr": "http://mlcommons.org/croissant/",
6
+ "dct": "http://purl.org/dc/terms/",
7
+ "prov": "http://www.w3.org/ns/prov#",
8
+ "rai": "http://mlcommons.org/croissant/RAI/",
9
+ "field": "cr:field",
10
+ "recordSet": "cr:recordSet",
11
+ "source": "cr:source",
12
+ "fileObject": "cr:fileObject",
13
+ "fileSet": "cr:fileSet",
14
+ "extract": "cr:extract",
15
+ "containedIn": "cr:containedIn",
16
+ "includes": "cr:includes",
17
+ "conformsTo": "dct:conformsTo",
18
+ "citeAs": "cr:citeAs"
19
+ },
20
+ "@type": "sc:Dataset",
21
+ "name": "Temporal Twins Benchmark",
22
+ "description": "Temporal Twins is a synthetic UPI-style transaction benchmark for temporal fraud detection. The collection contains oracle_calib, easy, medium, and hard matched-prefix benchmark slices across deterministic seeds 0, 1, 2, 3, and 4. Fraud labels are assigned through delayed temporal mechanisms rather than static per-transaction attributes, and matched fraud/benign twin examples are aligned at the same local prefix index to suppress static shortcuts while preserving order-sensitive temporal structure.",
23
+ "url": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins",
24
+ "license": "https://creativecommons.org/licenses/by/4.0/",
25
+ "isBasedOn": {
26
+ "@type": "sc:SoftwareSourceCode",
27
+ "name": "Temporal Twins benchmark code",
28
+ "url": "TODO_CODE_REPOSITORY_URL",
29
+ "license": "https://www.apache.org/licenses/LICENSE-2.0",
30
+ "identifier": "Apache-2.0"
31
+ },
32
+ "conformsTo": "http://mlcommons.org/croissant/1.1",
33
+ "citation": "TODO: Replace with the final Temporal Twins paper citation. Paper URL: TODO_PAPER_URL",
34
+ "citeAs": "Temporal Twins Benchmark (synthetic UPI-style temporal fraud benchmark), paper URL TODO_PAPER_URL, code repository TODO_CODE_REPOSITORY_URL.",
35
+ "creator": [
36
+ {
37
+ "@type": "sc:Organization",
38
+ "name": "Temporal Twins Benchmark Contributors"
39
+ }
40
+ ],
41
+ "dateCreated": "2026-05-04",
42
+ "version": "1.0.0",
43
+ "keywords": [
44
+ "synthetic financial transactions",
45
+ "UPI-style benchmark",
46
+ "temporal fraud detection",
47
+ "matched temporal twins",
48
+ "matched-prefix evaluation",
49
+ "sequence modeling",
50
+ "dynamic graph learning",
51
+ "reproducible benchmark"
52
+ ],
53
+ "distribution": [
54
+ {
55
+ "@id": "transactions-archive",
56
+ "@type": "cr:FileObject",
57
+ "name": "Transactions archive",
58
+ "description": "Hosted archive containing synthetic transaction files for oracle_calib, easy, medium, and hard across seeds 0 through 4.",
59
+ "contentUrl": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/tree/main/data",
60
+ "encodingFormat": "application/zip"
61
+ },
62
+ {
63
+ "@id": "matched-prefix-archive",
64
+ "@type": "cr:FileObject",
65
+ "name": "Matched-prefix examples archive",
66
+ "description": "Hosted release archive containing matched-prefix fraud/benign evaluation examples under release/data/*/seed_*/matched_pairs.parquet.",
67
+ "contentUrl": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins",
68
+ "encodingFormat": "application/zip"
69
+ },
70
+ {
71
+ "@id": "configs-archive",
72
+ "@type": "cr:FileObject",
73
+ "name": "Configs archive",
74
+ "description": "Hosted release archive containing benchmark configuration files under release/configs/*.yaml.",
75
+ "contentUrl": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins",
76
+ "encodingFormat": "application/zip"
77
+ },
78
+ {
79
+ "@id": "results-archive",
80
+ "@type": "cr:FileObject",
81
+ "name": "Results archive",
82
+ "description": "Hosted release archive containing the deterministic 5-seed paper-suite outputs under release/results/.",
83
+ "contentUrl": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins",
84
+ "encodingFormat": "application/zip"
85
+ },
86
+ {
87
+ "@id": "metadata-files",
88
+ "@type": "cr:FileSet",
89
+ "name": "Metadata files",
90
+ "description": "Metadata payload for the public release, including this Croissant file and companion notes.",
91
+ "containedIn": {
92
+ "@id": "results-archive"
93
+ },
94
+ "includes": "release/metadata/*"
95
+ },
96
+ {
97
+ "@id": "transactions-files",
98
+ "@type": "cr:FileSet",
99
+ "name": "Synthetic transactions parquet files",
100
+ "description": "Expected synthetic transaction files for benchmark modes oracle_calib, easy, medium, and hard across seeds 0 through 4.",
101
+ "containedIn": {
102
+ "@id": "transactions-archive"
103
+ },
104
+ "includes": "release/data/*/seed_*/transactions.parquet",
105
+ "encodingFormat": "application/x-parquet"
106
+ },
107
+ {
108
+ "@id": "matched-prefix-files",
109
+ "@type": "cr:FileSet",
110
+ "name": "Matched-prefix example parquet files",
111
+ "description": "Expected matched-prefix benchmark examples for the release. Each file contains fraud and benign twin examples evaluated at the same local prefix index.",
112
+ "containedIn": {
113
+ "@id": "matched-prefix-archive"
114
+ },
115
+ "includes": "release/data/*/seed_*/matched_pairs.parquet",
116
+ "encodingFormat": "application/x-parquet"
117
+ },
118
+ {
119
+ "@id": "config-files",
120
+ "@type": "cr:FileSet",
121
+ "name": "Benchmark config files",
122
+ "description": "YAML configuration files for the public release.",
123
+ "containedIn": {
124
+ "@id": "configs-archive"
125
+ },
126
+ "includes": "release/configs/*.yaml"
127
+ },
128
+ {
129
+ "@id": "paper-suite-runs-csv",
130
+ "@type": "cr:FileObject",
131
+ "name": "Per-run paper-suite results",
132
+ "description": "Per-run deterministic results for the final 5-seed paper-scale suite.",
133
+ "contentUrl": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/results/paper_suite_runs.csv",
134
+ "containedIn": {
135
+ "@id": "results-archive"
136
+ },
137
+ "encodingFormat": "text/csv"
138
+ },
139
+ {
140
+ "@id": "paper-suite-summary-csv",
141
+ "@type": "cr:FileObject",
142
+ "name": "Paper-suite summary results",
143
+ "description": "Mean and standard deviation summary of the deterministic 5-seed paper suite.",
144
+ "contentUrl": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/results/paper_suite_summary.csv",
145
+ "containedIn": {
146
+ "@id": "results-archive"
147
+ },
148
+ "encodingFormat": "text/csv"
149
+ },
150
+ {
151
+ "@id": "paper-suite-runtime-csv",
152
+ "@type": "cr:FileObject",
153
+ "name": "Paper-suite runtime summary",
154
+ "description": "Runtime and StaticGNN evaluation diagnostics for the final paper suite.",
155
+ "contentUrl": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/results/paper_suite_runtime.csv",
156
+ "containedIn": {
157
+ "@id": "results-archive"
158
+ },
159
+ "encodingFormat": "text/csv"
160
+ },
161
+ {
162
+ "@id": "paper-suite-failed-checks-csv",
163
+ "@type": "cr:FileObject",
164
+ "name": "Paper-suite failed gate checks",
165
+ "description": "Gate-check and advisory-check outcomes for each run in the final paper suite.",
166
+ "contentUrl": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/results/paper_suite_failed_checks.csv",
167
+ "containedIn": {
168
+ "@id": "results-archive"
169
+ },
170
+ "encodingFormat": "text/csv"
171
+ },
172
+ {
173
+ "@id": "croissant-file",
174
+ "@type": "cr:FileObject",
175
+ "name": "Temporal Twins Croissant metadata",
176
+ "description": "MLCommons Croissant 1.1 metadata for the full Temporal Twins benchmark collection.",
177
+ "contentUrl": "https://huggingface.co/datasets/temporal-twins-benchmark/temporal-twins/raw/main/metadata/temporal_twins_croissant.json",
178
+ "containedIn": {
179
+ "@id": "metadata-files"
180
+ },
181
+ "encodingFormat": "application/ld+json"
182
+ }
183
+ ],
184
+ "recordSet": [
185
+ {
186
+ "@id": "transactions",
187
+ "@type": "cr:RecordSet",
188
+ "name": "transactions",
189
+ "description": "Synthetic UPI-style transactions spanning oracle_calib, easy, medium, and hard, with deterministic seeds 0 through 4.",
190
+ "field": [
191
+ {
192
+ "@id": "transactions/sender_id",
193
+ "@type": "cr:Field",
194
+ "name": "sender_id",
195
+ "description": "Synthetic sender account identifier.",
196
+ "dataType": "sc:Text",
197
+ "source": {
198
+ "fileSet": {
199
+ "@id": "transactions-files"
200
+ },
201
+ "extract": {
202
+ "column": "sender_id"
203
+ }
204
+ }
205
+ },
206
+ {
207
+ "@id": "transactions/receiver_id",
208
+ "@type": "cr:Field",
209
+ "name": "receiver_id",
210
+ "description": "Synthetic receiver account identifier.",
211
+ "dataType": "sc:Text",
212
+ "source": {
213
+ "fileSet": {
214
+ "@id": "transactions-files"
215
+ },
216
+ "extract": {
217
+ "column": "receiver_id"
218
+ }
219
+ }
220
+ },
221
+ {
222
+ "@id": "transactions/timestamp",
223
+ "@type": "cr:Field",
224
+ "name": "timestamp",
225
+ "description": "Synthetic event timestamp used to order transactions within each sender history.",
226
+ "dataType": "sc:Number",
227
+ "source": {
228
+ "fileSet": {
229
+ "@id": "transactions-files"
230
+ },
231
+ "extract": {
232
+ "column": "timestamp"
233
+ }
234
+ }
235
+ },
236
+ {
237
+ "@id": "transactions/amount",
238
+ "@type": "cr:Field",
239
+ "name": "amount",
240
+ "description": "Synthetic transaction amount.",
241
+ "dataType": "sc:Number",
242
+ "source": {
243
+ "fileSet": {
244
+ "@id": "transactions-files"
245
+ },
246
+ "extract": {
247
+ "column": "amount"
248
+ }
249
+ }
250
+ },
251
+ {
252
+ "@id": "transactions/risk_score",
253
+ "@type": "cr:Field",
254
+ "name": "risk_score",
255
+ "description": "Synthetic noisy risk score emitted by the simulator's risk engine.",
256
+ "dataType": "sc:Number",
257
+ "source": {
258
+ "fileSet": {
259
+ "@id": "transactions-files"
260
+ },
261
+ "extract": {
262
+ "column": "risk_score"
263
+ }
264
+ }
265
+ },
266
+ {
267
+ "@id": "transactions/failed",
268
+ "@type": "cr:Field",
269
+ "name": "failed",
270
+ "description": "Indicator for whether the synthetic transaction attempt failed.",
271
+ "dataType": "sc:Boolean",
272
+ "source": {
273
+ "fileSet": {
274
+ "@id": "transactions-files"
275
+ },
276
+ "extract": {
277
+ "column": "failed"
278
+ }
279
+ }
280
+ },
281
+ {
282
+ "@id": "transactions/is_fraud",
283
+ "@type": "cr:Field",
284
+ "name": "is_fraud",
285
+ "description": "Delayed synthetic fraud label attached to specific transactions.",
286
+ "dataType": "sc:Boolean",
287
+ "source": {
288
+ "fileSet": {
289
+ "@id": "transactions-files"
290
+ },
291
+ "extract": {
292
+ "column": "is_fraud"
293
+ }
294
+ }
295
+ }
296
+ ]
297
+ },
298
+ {
299
+ "@id": "matched_prefix_examples",
300
+ "@type": "cr:RecordSet",
301
+ "name": "matched_prefix_examples",
302
+ "description": "Matched fraud and benign evaluation examples. Each benign twin is evaluated at the same local prefix index as the paired fraud twin, with matched static and prefix-level summaries. The release-facing field matched_local_event_idx is the matched prefix index and may correspond to the internal eval_local_event_idx column if files are exported directly from the current pipeline.",
303
+ "field": [
304
+ {
305
+ "@id": "matched_prefix_examples/twin_pair_id",
306
+ "@type": "cr:Field",
307
+ "name": "twin_pair_id",
308
+ "description": "Matched fraud/benign twin pair identifier.",
309
+ "dataType": "sc:Integer",
310
+ "source": {
311
+ "fileSet": {
312
+ "@id": "matched-prefix-files"
313
+ },
314
+ "extract": {
315
+ "column": "twin_pair_id"
316
+ }
317
+ }
318
+ },
319
+ {
320
+ "@id": "matched_prefix_examples/sender_id",
321
+ "@type": "cr:Field",
322
+ "name": "sender_id",
323
+ "description": "Sender evaluated at the matched prefix.",
324
+ "dataType": "sc:Text",
325
+ "source": {
326
+ "fileSet": {
327
+ "@id": "matched-prefix-files"
328
+ },
329
+ "extract": {
330
+ "column": "sender_id"
331
+ }
332
+ }
333
+ },
334
+ {
335
+ "@id": "matched_prefix_examples/matched_local_event_idx",
336
+ "@type": "cr:Field",
337
+ "name": "matched_local_event_idx",
338
+ "description": "Release-facing matched-prefix event index k used for both the fraud twin and its benign control.",
339
+ "dataType": "sc:Integer",
340
+ "source": {
341
+ "fileSet": {
342
+ "@id": "matched-prefix-files"
343
+ },
344
+ "extract": {
345
+ "column": "matched_local_event_idx"
346
+ }
347
+ }
348
+ },
349
+ {
350
+ "@id": "matched_prefix_examples/label",
351
+ "@type": "cr:Field",
352
+ "name": "label",
353
+ "description": "Binary matched-prefix label where 1 denotes the fraud twin example and 0 denotes the benign matched control.",
354
+ "dataType": "sc:Boolean",
355
+ "source": {
356
+ "fileSet": {
357
+ "@id": "matched-prefix-files"
358
+ },
359
+ "extract": {
360
+ "column": "label"
361
+ }
362
+ }
363
+ },
364
+ {
365
+ "@id": "matched_prefix_examples/benchmark_mode",
366
+ "@type": "cr:Field",
367
+ "name": "benchmark_mode",
368
+ "description": "Benchmark mode identifier, e.g. temporal_twins_oracle_calib or temporal_twins.",
369
+ "dataType": "sc:Text",
370
+ "source": {
371
+ "fileSet": {
372
+ "@id": "matched-prefix-files"
373
+ },
374
+ "extract": {
375
+ "column": "benchmark_mode"
376
+ }
377
+ }
378
+ },
379
+ {
380
+ "@id": "matched_prefix_examples/difficulty",
381
+ "@type": "cr:Field",
382
+ "name": "difficulty",
383
+ "description": "Difficulty slice within the release: oracle_calib, easy, medium, or hard.",
384
+ "dataType": "sc:Text",
385
+ "source": {
386
+ "fileSet": {
387
+ "@id": "matched-prefix-files"
388
+ },
389
+ "extract": {
390
+ "column": "difficulty"
391
+ }
392
+ }
393
+ },
394
+ {
395
+ "@id": "matched_prefix_examples/seed",
396
+ "@type": "cr:Field",
397
+ "name": "seed",
398
+ "description": "Deterministic benchmark seed in the final paper-scale suite.",
399
+ "dataType": "sc:Integer",
400
+ "source": {
401
+ "fileSet": {
402
+ "@id": "matched-prefix-files"
403
+ },
404
+ "extract": {
405
+ "column": "seed"
406
+ }
407
+ }
408
+ }
409
+ ]
410
+ },
411
+ {
412
+ "@id": "audit_columns",
413
+ "@type": "cr:RecordSet",
414
+ "name": "audit_columns",
415
+ "description": "Audit and probe support columns carried by the synthetic generator for analysis, oracle-style scoring, and benchmark validation. These columns are not intended for ordinary model training and should be excluded from learned baseline inputs in benchmark evaluations.",
416
+ "field": [
417
+ {
418
+ "@id": "audit_columns/twin_role",
419
+ "@type": "cr:Field",
420
+ "name": "twin_role",
421
+ "description": "Twin role label such as fraud, benign, or background; excluded from ordinary model features.",
422
+ "dataType": "sc:Text",
423
+ "source": {
424
+ "fileSet": {
425
+ "@id": "transactions-files"
426
+ },
427
+ "extract": {
428
+ "column": "twin_role"
429
+ }
430
+ }
431
+ },
432
+ {
433
+ "@id": "audit_columns/template_id",
434
+ "@type": "cr:Field",
435
+ "name": "template_id",
436
+ "description": "Identifier for the matched temporal template used to construct a twin pair; excluded from ordinary model features.",
437
+ "dataType": "sc:Integer",
438
+ "source": {
439
+ "fileSet": {
440
+ "@id": "transactions-files"
441
+ },
442
+ "extract": {
443
+ "column": "template_id"
444
+ }
445
+ }
446
+ },
447
+ {
448
+ "@id": "audit_columns/motif_hit_count",
449
+ "@type": "cr:Field",
450
+ "name": "motif_hit_count",
451
+ "description": "Count of motif hits in the generator trace; exposed only for audit or probe logic, not learned baselines.",
452
+ "dataType": "sc:Integer",
453
+ "source": {
454
+ "fileSet": {
455
+ "@id": "transactions-files"
456
+ },
457
+ "extract": {
458
+ "column": "motif_hit_count"
459
+ }
460
+ }
461
+ },
462
+ {
463
+ "@id": "audit_columns/motif_source",
464
+ "@type": "cr:Field",
465
+ "name": "motif_source",
466
+ "description": "Generator-side motif provenance label; excluded from ordinary model features.",
467
+ "dataType": "sc:Text",
468
+ "source": {
469
+ "fileSet": {
470
+ "@id": "transactions-files"
471
+ },
472
+ "extract": {
473
+ "column": "motif_source"
474
+ }
475
+ }
476
+ },
477
+ {
478
+ "@id": "audit_columns/trigger_event_idx",
479
+ "@type": "cr:Field",
480
+ "name": "trigger_event_idx",
481
+ "description": "Internal trigger event index for delayed fraud assignment; excluded from ordinary model features.",
482
+ "dataType": "sc:Integer",
483
+ "source": {
484
+ "fileSet": {
485
+ "@id": "transactions-files"
486
+ },
487
+ "extract": {
488
+ "column": "trigger_event_idx"
489
+ }
490
+ }
491
+ },
492
+ {
493
+ "@id": "audit_columns/label_event_idx",
494
+ "@type": "cr:Field",
495
+ "name": "label_event_idx",
496
+ "description": "Internal event index at which the delayed fraud label is attached; excluded from ordinary model features.",
497
+ "dataType": "sc:Integer",
498
+ "source": {
499
+ "fileSet": {
500
+ "@id": "transactions-files"
501
+ },
502
+ "extract": {
503
+ "column": "label_event_idx"
504
+ }
505
+ }
506
+ },
507
+ {
508
+ "@id": "audit_columns/label_delay",
509
+ "@type": "cr:Field",
510
+ "name": "label_delay",
511
+ "description": "Internal delay between trigger and labeled event; excluded from ordinary model features.",
512
+ "dataType": "sc:Integer",
513
+ "source": {
514
+ "fileSet": {
515
+ "@id": "transactions-files"
516
+ },
517
+ "extract": {
518
+ "column": "label_delay"
519
+ }
520
+ }
521
+ },
522
+ {
523
+ "@id": "audit_columns/fraud_source",
524
+ "@type": "cr:Field",
525
+ "name": "fraud_source",
526
+ "description": "Internal fraud-source annotation such as motif or fallback; excluded from ordinary model features.",
527
+ "dataType": "sc:Text",
528
+ "source": {
529
+ "fileSet": {
530
+ "@id": "transactions-files"
531
+ },
532
+ "extract": {
533
+ "column": "fraud_source"
534
+ }
535
+ }
536
+ },
537
+ {
538
+ "@id": "audit_columns/dynamic_fraud_state",
539
+ "@type": "cr:Field",
540
+ "name": "dynamic_fraud_state",
541
+ "description": "Latent generator-side fraud-state variable used for mechanistic analysis; excluded from ordinary model features.",
542
+ "dataType": "sc:Number",
543
+ "source": {
544
+ "fileSet": {
545
+ "@id": "transactions-files"
546
+ },
547
+ "extract": {
548
+ "column": "dynamic_fraud_state"
549
+ }
550
+ }
551
+ }
552
+ ]
553
+ },
554
+ {
555
+ "@id": "paper_suite_summary_results",
556
+ "@type": "cr:RecordSet",
557
+ "name": "paper_suite_summary_results",
558
+ "description": "Deterministic 5-seed summary results for the final paper-scale Temporal Twins suite.",
559
+ "field": [
560
+ {
561
+ "@id": "paper_suite_summary_results/benchmark_group",
562
+ "@type": "cr:Field",
563
+ "name": "benchmark_group",
564
+ "description": "Benchmark slice summarized in the row, e.g. oracle_calib, easy, medium, or hard.",
565
+ "dataType": "sc:Text",
566
+ "source": {
567
+ "fileObject": {
568
+ "@id": "paper-suite-summary-csv"
569
+ },
570
+ "extract": {
571
+ "column": "benchmark_group"
572
+ }
573
+ }
574
+ },
575
+ {
576
+ "@id": "paper_suite_summary_results/matched_eval_pairs_mean",
577
+ "@type": "cr:Field",
578
+ "name": "matched_eval_pairs_mean",
579
+ "description": "Mean number of matched-prefix evaluation pairs across seeds.",
580
+ "dataType": "sc:Number",
581
+ "source": {
582
+ "fileObject": {
583
+ "@id": "paper-suite-summary-csv"
584
+ },
585
+ "extract": {
586
+ "column": "matched_eval_pairs_mean"
587
+ }
588
+ }
589
+ },
590
+ {
591
+ "@id": "paper_suite_summary_results/static_agg_auc_mean",
592
+ "@type": "cr:Field",
593
+ "name": "static_agg_auc_mean",
594
+ "description": "Mean ROC-AUC of the static aggregate shortcut audit.",
595
+ "dataType": "sc:Number",
596
+ "source": {
597
+ "fileObject": {
598
+ "@id": "paper-suite-summary-csv"
599
+ },
600
+ "extract": {
601
+ "column": "static_agg_auc_mean"
602
+ }
603
+ }
604
+ },
605
+ {
606
+ "@id": "paper_suite_summary_results/audit_roc_auc_mean",
607
+ "@type": "cr:Field",
608
+ "name": "audit_roc_auc_mean",
609
+ "description": "Mean oracle or probe ROC-AUC depending on benchmark mode.",
610
+ "dataType": "sc:Number",
611
+ "source": {
612
+ "fileObject": {
613
+ "@id": "paper-suite-summary-csv"
614
+ },
615
+ "extract": {
616
+ "column": "audit_roc_auc_mean"
617
+ }
618
+ }
619
+ },
620
+ {
621
+ "@id": "paper_suite_summary_results/raw_roc_auc_mean",
622
+ "@type": "cr:Field",
623
+ "name": "raw_roc_auc_mean",
624
+ "description": "Mean raw motif oracle or probe ROC-AUC depending on benchmark mode.",
625
+ "dataType": "sc:Number",
626
+ "source": {
627
+ "fileObject": {
628
+ "@id": "paper-suite-summary-csv"
629
+ },
630
+ "extract": {
631
+ "column": "raw_roc_auc_mean"
632
+ }
633
+ }
634
+ },
635
+ {
636
+ "@id": "paper_suite_summary_results/xgb_roc_auc_mean",
637
+ "@type": "cr:Field",
638
+ "name": "xgb_roc_auc_mean",
639
+ "description": "Mean XGBoost ROC-AUC across seeds.",
640
+ "dataType": "sc:Number",
641
+ "source": {
642
+ "fileObject": {
643
+ "@id": "paper-suite-summary-csv"
644
+ },
645
+ "extract": {
646
+ "column": "xgb_roc_auc_mean"
647
+ }
648
+ }
649
+ },
650
+ {
651
+ "@id": "paper_suite_summary_results/static_gnn_roc_auc_mean",
652
+ "@type": "cr:Field",
653
+ "name": "static_gnn_roc_auc_mean",
654
+ "description": "Mean StaticGNN ROC-AUC across seeds.",
655
+ "dataType": "sc:Number",
656
+ "source": {
657
+ "fileObject": {
658
+ "@id": "paper-suite-summary-csv"
659
+ },
660
+ "extract": {
661
+ "column": "static_gnn_roc_auc_mean"
662
+ }
663
+ }
664
+ },
665
+ {
666
+ "@id": "paper_suite_summary_results/seqgru_clean_roc_auc_mean",
667
+ "@type": "cr:Field",
668
+ "name": "seqgru_clean_roc_auc_mean",
669
+ "description": "Mean clean SeqGRU ROC-AUC across seeds.",
670
+ "dataType": "sc:Number",
671
+ "source": {
672
+ "fileObject": {
673
+ "@id": "paper-suite-summary-csv"
674
+ },
675
+ "extract": {
676
+ "column": "seqgru_clean_roc_auc_mean"
677
+ }
678
+ }
679
+ },
680
+ {
681
+ "@id": "paper_suite_summary_results/seqgru_shuffle_delta_mean",
682
+ "@type": "cr:Field",
683
+ "name": "seqgru_shuffle_delta_mean",
684
+ "description": "Mean change in SeqGRU ROC-AUC under shuffled event order.",
685
+ "dataType": "sc:Number",
686
+ "source": {
687
+ "fileObject": {
688
+ "@id": "paper-suite-summary-csv"
689
+ },
690
+ "extract": {
691
+ "column": "seqgru_shuffle_delta_mean"
692
+ }
693
+ }
694
+ },
695
+ {
696
+ "@id": "paper_suite_summary_results/tgn_clean_roc_auc_mean",
697
+ "@type": "cr:Field",
698
+ "name": "tgn_clean_roc_auc_mean",
699
+ "description": "Mean TGN ROC-AUC across seeds.",
700
+ "dataType": "sc:Number",
701
+ "source": {
702
+ "fileObject": {
703
+ "@id": "paper-suite-summary-csv"
704
+ },
705
+ "extract": {
706
+ "column": "tgn_clean_roc_auc_mean"
707
+ }
708
+ }
709
+ },
710
+ {
711
+ "@id": "paper_suite_summary_results/tgat_clean_roc_auc_mean",
712
+ "@type": "cr:Field",
713
+ "name": "tgat_clean_roc_auc_mean",
714
+ "description": "Mean TGAT ROC-AUC across seeds.",
715
+ "dataType": "sc:Number",
716
+ "source": {
717
+ "fileObject": {
718
+ "@id": "paper-suite-summary-csv"
719
+ },
720
+ "extract": {
721
+ "column": "tgat_clean_roc_auc_mean"
722
+ }
723
+ }
724
+ },
725
+ {
726
+ "@id": "paper_suite_summary_results/dyrep_clean_roc_auc_mean",
727
+ "@type": "cr:Field",
728
+ "name": "dyrep_clean_roc_auc_mean",
729
+ "description": "Mean DyRep ROC-AUC across seeds.",
730
+ "dataType": "sc:Number",
731
+ "source": {
732
+ "fileObject": {
733
+ "@id": "paper-suite-summary-csv"
734
+ },
735
+ "extract": {
736
+ "column": "dyrep_clean_roc_auc_mean"
737
+ }
738
+ }
739
+ },
740
+ {
741
+ "@id": "paper_suite_summary_results/jodie_clean_roc_auc_mean",
742
+ "@type": "cr:Field",
743
+ "name": "jodie_clean_roc_auc_mean",
744
+ "description": "Mean JODIE ROC-AUC across seeds.",
745
+ "dataType": "sc:Number",
746
+ "source": {
747
+ "fileObject": {
748
+ "@id": "paper-suite-summary-csv"
749
+ },
750
+ "extract": {
751
+ "column": "jodie_clean_roc_auc_mean"
752
+ }
753
+ }
754
+ }
755
+ ]
756
+ }
757
+ ],
758
+ "rai:dataLimitations": [
759
+ "Temporal Twins is fully synthetic and is not representative of real UPI fraud prevalence, transaction mix, or institutional controls.",
760
+ "The benchmark is designed to isolate temporal-order reasoning under matched static controls rather than to reproduce a production fraud environment.",
761
+ "Standard-mode probe scores are informative benchmark probes, not upper bounds on real-world fraud detectability."
762
+ ],
763
+ "rai:dataBiases": [
764
+ "Behavioral patterns are simulator-defined and reflect the assumptions of the Temporal Twins generator rather than observed user behavior.",
765
+ "Difficulty slices intentionally reshape motif strength, noise, delay, and adversarial perturbations, so conclusions should be interpreted as benchmark-relative rather than population-representative."
766
+ ],
767
+ "rai:personalSensitiveInformation": "None. The dataset contains no real UPI data, no real users, no real bank accounts, no real transactions, no personal financial records, and no protected demographic attributes.",
768
+ "rai:dataUseCases": [
769
+ "Intended for temporal machine learning benchmark research, including sequence models, dynamic graph models, matched-control evaluation, and shortcut auditing.",
770
+ "Suitable for studying whether a model uses causal temporal order rather than static transaction summaries."
771
+ ],
772
+ "rai:dataSocialImpact": [
773
+ "Positive use may include more rigorous evaluation of temporal fraud-detection methods under matched static controls.",
774
+ "Potential misuse includes treating synthetic behavior as if it were real user behavior or using the dataset to justify deployment decisions without external validation on real, appropriately governed data."
775
+ ],
776
+ "rai:hasSyntheticData": true,
777
+ "prov:wasGeneratedBy": {
778
+ "@type": "prov:Activity",
779
+ "name": "Temporal Twins synthetic UPI transaction generator",
780
+ "description": "Synthetic benchmark generation for oracle_calib, easy, medium, and hard using deterministic seeds [0, 1, 2, 3, 4], num_users=350, simulation_days=45, fast_mode=false, and n_checkpoints=8. The generator emits matched fraud/benign twins evaluated at matched local prefix indices and preserves paper-suite shortcut audits and summary results.",
781
+ "prov:used": [
782
+ {
783
+ "@type": "prov:Entity",
784
+ "name": "Temporal Twins benchmark code repository",
785
+ "url": "TODO_CODE_REPOSITORY_URL",
786
+ "license": "https://www.apache.org/licenses/LICENSE-2.0",
787
+ "identifier": "Apache-2.0"
788
+ },
789
+ {
790
+ "@type": "prov:Entity",
791
+ "name": "Temporal Twins paper",
792
+ "url": "TODO_PAPER_URL"
793
+ }
794
+ ]
795
+ }
796
+ }
data/.DS_Store ADDED
Binary file (10.2 kB). View file
 
data/README_GENERATION.md ADDED
@@ -0,0 +1,120 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Generating Release Data Files
2
+
3
+ The repository currently includes the **results** of the final paper suite, but it does **not** include pre-exported per-seed release files under `release/data/`. This document explains how to generate them using the existing Temporal Twins benchmark code without changing generator logic, labels, matched-prefix construction, or model logic.
4
+
5
+ ## Expected Outputs Per Seed
6
+
7
+ Each directory `release/data/<mode>/seed_<seed>/` is expected to contain:
8
+
9
+ - `transactions.parquet`
10
+ - `matched_pairs.parquet`
11
+ - `audit_summary.csv`
12
+ - `schema.json`
13
+ - `config.yaml`
14
+
15
+ Where:
16
+
17
+ - `<mode>` is one of `oracle_calib`, `easy`, `medium`, `hard`
18
+ - `<seed>` is one of `0`, `1`, `2`, `3`, `4`
19
+
20
+ ## Benchmark Mapping
21
+
22
+ - `oracle_calib` uses `benchmark_mode = "temporal_twins_oracle_calib"` and `difficulty = "easy"`
23
+ - `easy` uses `benchmark_mode = "temporal_twins"` and `difficulty = "easy"`
24
+ - `medium` uses `benchmark_mode = "temporal_twins"` and `difficulty = "medium"`
25
+ - `hard` uses `benchmark_mode = "temporal_twins"` and `difficulty = "hard"`
26
+
27
+ ## Exact Export Command
28
+
29
+ Run this command from the repository root:
30
+
31
+ ```bash
32
+ PYTHONPATH=. python3 - <<'PY'
33
+ from pathlib import Path
34
+ import json
35
+ import pandas as pd
36
+ import yaml
37
+
38
+ from src.core.config_loader import load_config
39
+ from experiments.run_all import (
40
+ build_matched_control_tables,
41
+ generate_single_difficulty,
42
+ report_matched_control_audits,
43
+ set_global_determinism,
44
+ )
45
+
46
+ release_root = Path("release/data")
47
+ seeds = [0, 1, 2, 3, 4]
48
+ mode_specs = [
49
+ ("oracle_calib", "temporal_twins_oracle_calib", "easy"),
50
+ ("easy", "temporal_twins", "easy"),
51
+ ("medium", "temporal_twins", "medium"),
52
+ ("hard", "temporal_twins", "hard"),
53
+ ]
54
+
55
+ base_cfg = load_config("config/default.yaml")
56
+ base_cfg.num_users = 350
57
+ base_cfg.simulation_days = 45
58
+
59
+ for release_mode, benchmark_mode, difficulty in mode_specs:
60
+ for seed in seeds:
61
+ cfg = base_cfg.model_copy(deep=True)
62
+ cfg.benchmark_mode = benchmark_mode
63
+ cfg.random_seed = seed
64
+ set_global_determinism(seed)
65
+
66
+ df = generate_single_difficulty(
67
+ cfg,
68
+ difficulty=difficulty,
69
+ seed=seed,
70
+ benchmark_mode=benchmark_mode,
71
+ )
72
+ matched_examples, pair_rows, pair_counts = build_matched_control_tables(df)
73
+ audit = report_matched_control_audits(matched_examples, pair_rows, pair_counts)
74
+
75
+ out_dir = release_root / release_mode / f"seed_{seed}"
76
+ out_dir.mkdir(parents=True, exist_ok=True)
77
+
78
+ matched_export = matched_examples.rename(
79
+ columns={"eval_local_event_idx": "matched_local_event_idx"}
80
+ ).copy()
81
+ matched_export["benchmark_mode"] = benchmark_mode
82
+ matched_export["difficulty"] = release_mode
83
+ matched_export["seed"] = seed
84
+
85
+ df.to_parquet(out_dir / "transactions.parquet", index=False)
86
+ matched_export.to_parquet(out_dir / "matched_pairs.parquet", index=False)
87
+ pd.DataFrame([audit]).to_csv(out_dir / "audit_summary.csv", index=False)
88
+
89
+ schema = {
90
+ "transactions_columns": {k: str(v) for k, v in df.dtypes.items()},
91
+ "matched_pairs_columns": {k: str(v) for k, v in matched_export.dtypes.items()},
92
+ "files": [
93
+ "transactions.parquet",
94
+ "matched_pairs.parquet",
95
+ "audit_summary.csv",
96
+ "schema.json",
97
+ "config.yaml",
98
+ ],
99
+ }
100
+ (out_dir / "schema.json").write_text(json.dumps(schema, indent=2) + "\\n")
101
+ (out_dir / "config.yaml").write_text(
102
+ yaml.safe_dump(
103
+ {
104
+ **cfg.model_dump(),
105
+ "benchmark_mode": benchmark_mode,
106
+ "difficulty": difficulty,
107
+ "release_mode": release_mode,
108
+ "seed": seed,
109
+ "fast_mode": False,
110
+ "n_checkpoints": 8,
111
+ },
112
+ sort_keys=False,
113
+ )
114
+ )
115
+ PY
116
+ ```
117
+
118
+ ## Paper Result Reproduction
119
+
120
+ After generating the release data files, the final paper-suite metrics can be reproduced from the benchmark runner with the frozen deterministic settings and the same `num_users`, `simulation_days`, `seeds`, and `n_checkpoints` recorded in `release/results/paper_suite_meta.json`.
data/_export_summary.csv ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ release_mode,seed,transactions,matched_examples,matched_pairs,audit_examples
2
+ oracle_calib,0,35408,2240,1120,2240
3
+ oracle_calib,1,31994,2106,1053,2106
4
+ oracle_calib,2,35922,2390,1195,2390
5
+ oracle_calib,3,32228,2222,1111,2222
6
+ oracle_calib,4,32108,2276,1138,2276
7
+ easy,0,46386,3398,1699,3398
8
+ easy,1,40462,3132,1566,3132
9
+ easy,2,44958,3558,1779,3558
10
+ easy,3,46312,3374,1687,3374
11
+ easy,4,41482,3296,1648,3296
12
+ medium,0,77692,3184,1592,3184
13
+ medium,1,76870,3168,1584,3168
14
+ medium,2,80150,3174,1587,3174
15
+ medium,3,80172,3148,1574,3148
16
+ medium,4,78456,3144,1572,3144
17
+ hard,0,81978,2600,1300,2600
18
+ hard,1,97936,2656,1328,2656
19
+ hard,2,86358,2624,1312,2624
20
+ hard,3,76070,2614,1307,2614
21
+ hard,4,76406,2618,1309,2618
data/easy/.DS_Store ADDED
Binary file (8.2 kB). View file
 
data/easy/seed_0/audit_summary.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ pair_total_txn_count_diff_mean,pair_total_txn_count_diff_max,auc_total_txn_count,auc_local_event_idx,auc_prefix_txn_count,auc_timestamp,auc_account_age,auc_active_age,fraud_label_event_idx_mean,fraud_label_event_idx_max,benign_eval_event_idx_mean,benign_eval_event_idx_max,pair_event_idx_diff_mean,pair_event_idx_diff_max,pair_active_age_diff_mean,pair_active_age_diff_max,pair_timestamp_diff_mean,pair_timestamp_diff_max,benign_motif_hit_rate,benign_motif_hit_pairs,matched_control_examples,matched_control_pair_events
2
+ 0.0,0.0,0.5,0.49999999999999994,0.49999999999999994,0.5,0.5,0.5,40.72630959387875,104.0,40.72630959387875,104.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0,3398,1699
data/easy/seed_0/config.yaml ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_users: 350
2
+ simulation_days: 45
3
+ fraud_ratio: 0.05
4
+ benchmark_mode: temporal_twins
5
+ user_params:
6
+ lambda_mean: 5.0
7
+ lambda_std: 1.0
8
+ mu_mean: 7.5
9
+ mu_std: 1.0
10
+ sigma_mean: 0.5
11
+ sigma_std: 0.2
12
+ upi_limits:
13
+ max_txn_amount: 100000.0
14
+ daily_limit: 100000.0
15
+ risk_model:
16
+ weights:
17
+ amount_ratio: 1.0
18
+ daily_ratio: 0.8
19
+ velocity: 1.2
20
+ time_anomaly: 0.6
21
+ graph_anomaly: 1.0
22
+ retry: 0.8
23
+ kyc: 0.5
24
+ user_risk: 0.8
25
+ random_seed: 0
26
+ difficulty: easy
27
+ release_mode: easy
28
+ seed: 0
29
+ fast_mode: false
30
+ n_checkpoints: 8
data/easy/seed_0/matched_pairs.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d08d99bb61bfc154f2020f20458f4ba34084737ca12c191784d00e4c5f9a968b
3
+ size 81815
data/easy/seed_0/schema.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transactions_columns": {
3
+ "txn_id": "int32",
4
+ "sender_id": "int64",
5
+ "receiver_id": "int64",
6
+ "amount": "float32",
7
+ "timestamp": "float32",
8
+ "txn_type": "int8",
9
+ "is_fraud": "int8",
10
+ "fraud_type": "str",
11
+ "is_retry": "int8",
12
+ "risk_score": "float32",
13
+ "fail_prob": "float32",
14
+ "failed": "int8",
15
+ "twin_pair_id": "int64",
16
+ "template_id": "int64",
17
+ "twin_role": "str",
18
+ "twin_label": "int8",
19
+ "motif_source": "int8",
20
+ "motif_chain_state": "float32",
21
+ "motif_strength": "float32",
22
+ "dynamic_fraud_state": "float32",
23
+ "fraud_source": "str",
24
+ "motif_hit_count": "int32",
25
+ "trigger_event_idx": "int32",
26
+ "label_event_idx": "int32",
27
+ "label_delay": "int32",
28
+ "is_fallback_label": "int8",
29
+ "risk_noisy": "float32",
30
+ "neighbor_score": "float32",
31
+ "pair_freq": "float32",
32
+ "txn_count_10": "float32",
33
+ "amount_sum_10": "float32"
34
+ },
35
+ "matched_pairs_columns": {
36
+ "pair_event_id": "int64",
37
+ "twin_pair_id": "int64",
38
+ "template_id": "int64",
39
+ "matched_local_event_idx": "int64",
40
+ "prefix_txn_count": "int64",
41
+ "sender_id": "int64",
42
+ "label": "int64",
43
+ "twin_role": "str",
44
+ "matched_sender_id": "int64",
45
+ "total_txn_count": "int64",
46
+ "eval_timestamp": "float64",
47
+ "account_age": "float64",
48
+ "active_age": "float64",
49
+ "benchmark_mode": "str",
50
+ "difficulty": "str",
51
+ "seed": "int64"
52
+ },
53
+ "audit_summary_columns": {
54
+ "pair_total_txn_count_diff_mean": "float64",
55
+ "pair_total_txn_count_diff_max": "float64",
56
+ "auc_total_txn_count": "float64",
57
+ "auc_local_event_idx": "float64",
58
+ "auc_prefix_txn_count": "float64",
59
+ "auc_timestamp": "float64",
60
+ "auc_account_age": "float64",
61
+ "auc_active_age": "float64",
62
+ "fraud_label_event_idx_mean": "float64",
63
+ "fraud_label_event_idx_max": "float64",
64
+ "benign_eval_event_idx_mean": "float64",
65
+ "benign_eval_event_idx_max": "float64",
66
+ "pair_event_idx_diff_mean": "float64",
67
+ "pair_event_idx_diff_max": "float64",
68
+ "pair_active_age_diff_mean": "float64",
69
+ "pair_active_age_diff_max": "float64",
70
+ "pair_timestamp_diff_mean": "float64",
71
+ "pair_timestamp_diff_max": "float64",
72
+ "benign_motif_hit_rate": "float64",
73
+ "benign_motif_hit_pairs": "int64",
74
+ "matched_control_examples": "int64",
75
+ "matched_control_pair_events": "int64"
76
+ },
77
+ "files": [
78
+ "transactions.parquet",
79
+ "matched_pairs.parquet",
80
+ "audit_summary.csv",
81
+ "schema.json",
82
+ "config.yaml"
83
+ ]
84
+ }
data/easy/seed_0/transactions.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8024a68ece6179d8d3af63ffd66eb8a444c79737760049dda816a03e4a25dbd7
3
+ size 1332635
data/easy/seed_1/audit_summary.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ pair_total_txn_count_diff_mean,pair_total_txn_count_diff_max,auc_total_txn_count,auc_local_event_idx,auc_prefix_txn_count,auc_timestamp,auc_account_age,auc_active_age,fraud_label_event_idx_mean,fraud_label_event_idx_max,benign_eval_event_idx_mean,benign_eval_event_idx_max,pair_event_idx_diff_mean,pair_event_idx_diff_max,pair_active_age_diff_mean,pair_active_age_diff_max,pair_timestamp_diff_mean,pair_timestamp_diff_max,benign_motif_hit_rate,benign_motif_hit_pairs,matched_control_examples,matched_control_pair_events
2
+ 0.0,0.0,0.5,0.49999999999999994,0.49999999999999994,0.5000000000000001,0.5,0.5,36.748403575989784,100.0,36.748403575989784,100.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0,3132,1566
data/easy/seed_1/config.yaml ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_users: 350
2
+ simulation_days: 45
3
+ fraud_ratio: 0.05
4
+ benchmark_mode: temporal_twins
5
+ user_params:
6
+ lambda_mean: 5.0
7
+ lambda_std: 1.0
8
+ mu_mean: 7.5
9
+ mu_std: 1.0
10
+ sigma_mean: 0.5
11
+ sigma_std: 0.2
12
+ upi_limits:
13
+ max_txn_amount: 100000.0
14
+ daily_limit: 100000.0
15
+ risk_model:
16
+ weights:
17
+ amount_ratio: 1.0
18
+ daily_ratio: 0.8
19
+ velocity: 1.2
20
+ time_anomaly: 0.6
21
+ graph_anomaly: 1.0
22
+ retry: 0.8
23
+ kyc: 0.5
24
+ user_risk: 0.8
25
+ random_seed: 1
26
+ difficulty: easy
27
+ release_mode: easy
28
+ seed: 1
29
+ fast_mode: false
30
+ n_checkpoints: 8
data/easy/seed_1/matched_pairs.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7503b30b1b597adc9507fb10361d8ef0c0ca6e5add89c502843caba3a0aa9e74
3
+ size 77643
data/easy/seed_1/schema.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transactions_columns": {
3
+ "txn_id": "int32",
4
+ "sender_id": "int64",
5
+ "receiver_id": "int64",
6
+ "amount": "float32",
7
+ "timestamp": "float32",
8
+ "txn_type": "int8",
9
+ "is_fraud": "int8",
10
+ "fraud_type": "str",
11
+ "is_retry": "int8",
12
+ "risk_score": "float32",
13
+ "fail_prob": "float32",
14
+ "failed": "int8",
15
+ "twin_pair_id": "int64",
16
+ "template_id": "int64",
17
+ "twin_role": "str",
18
+ "twin_label": "int8",
19
+ "motif_source": "int8",
20
+ "motif_chain_state": "float32",
21
+ "motif_strength": "float32",
22
+ "dynamic_fraud_state": "float32",
23
+ "fraud_source": "str",
24
+ "motif_hit_count": "int32",
25
+ "trigger_event_idx": "int32",
26
+ "label_event_idx": "int32",
27
+ "label_delay": "int32",
28
+ "is_fallback_label": "int8",
29
+ "risk_noisy": "float32",
30
+ "neighbor_score": "float32",
31
+ "pair_freq": "float32",
32
+ "txn_count_10": "float32",
33
+ "amount_sum_10": "float32"
34
+ },
35
+ "matched_pairs_columns": {
36
+ "pair_event_id": "int64",
37
+ "twin_pair_id": "int64",
38
+ "template_id": "int64",
39
+ "matched_local_event_idx": "int64",
40
+ "prefix_txn_count": "int64",
41
+ "sender_id": "int64",
42
+ "label": "int64",
43
+ "twin_role": "str",
44
+ "matched_sender_id": "int64",
45
+ "total_txn_count": "int64",
46
+ "eval_timestamp": "float64",
47
+ "account_age": "float64",
48
+ "active_age": "float64",
49
+ "benchmark_mode": "str",
50
+ "difficulty": "str",
51
+ "seed": "int64"
52
+ },
53
+ "audit_summary_columns": {
54
+ "pair_total_txn_count_diff_mean": "float64",
55
+ "pair_total_txn_count_diff_max": "float64",
56
+ "auc_total_txn_count": "float64",
57
+ "auc_local_event_idx": "float64",
58
+ "auc_prefix_txn_count": "float64",
59
+ "auc_timestamp": "float64",
60
+ "auc_account_age": "float64",
61
+ "auc_active_age": "float64",
62
+ "fraud_label_event_idx_mean": "float64",
63
+ "fraud_label_event_idx_max": "float64",
64
+ "benign_eval_event_idx_mean": "float64",
65
+ "benign_eval_event_idx_max": "float64",
66
+ "pair_event_idx_diff_mean": "float64",
67
+ "pair_event_idx_diff_max": "float64",
68
+ "pair_active_age_diff_mean": "float64",
69
+ "pair_active_age_diff_max": "float64",
70
+ "pair_timestamp_diff_mean": "float64",
71
+ "pair_timestamp_diff_max": "float64",
72
+ "benign_motif_hit_rate": "float64",
73
+ "benign_motif_hit_pairs": "int64",
74
+ "matched_control_examples": "int64",
75
+ "matched_control_pair_events": "int64"
76
+ },
77
+ "files": [
78
+ "transactions.parquet",
79
+ "matched_pairs.parquet",
80
+ "audit_summary.csv",
81
+ "schema.json",
82
+ "config.yaml"
83
+ ]
84
+ }
data/easy/seed_1/transactions.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:22f5283274a62e56a22458610709081d8b1bbf6e4793b6363c3ca9e518dc7f3e
3
+ size 1229511
data/easy/seed_2/audit_summary.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ pair_total_txn_count_diff_mean,pair_total_txn_count_diff_max,auc_total_txn_count,auc_local_event_idx,auc_prefix_txn_count,auc_timestamp,auc_account_age,auc_active_age,fraud_label_event_idx_mean,fraud_label_event_idx_max,benign_eval_event_idx_mean,benign_eval_event_idx_max,pair_event_idx_diff_mean,pair_event_idx_diff_max,pair_active_age_diff_mean,pair_active_age_diff_max,pair_timestamp_diff_mean,pair_timestamp_diff_max,benign_motif_hit_rate,benign_motif_hit_pairs,matched_control_examples,matched_control_pair_events
2
+ 0.0,0.0,0.5,0.5,0.5,0.5,0.5,0.5,41.19336706014615,107.0,41.19336706014615,107.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0,3558,1779
data/easy/seed_2/config.yaml ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_users: 350
2
+ simulation_days: 45
3
+ fraud_ratio: 0.05
4
+ benchmark_mode: temporal_twins
5
+ user_params:
6
+ lambda_mean: 5.0
7
+ lambda_std: 1.0
8
+ mu_mean: 7.5
9
+ mu_std: 1.0
10
+ sigma_mean: 0.5
11
+ sigma_std: 0.2
12
+ upi_limits:
13
+ max_txn_amount: 100000.0
14
+ daily_limit: 100000.0
15
+ risk_model:
16
+ weights:
17
+ amount_ratio: 1.0
18
+ daily_ratio: 0.8
19
+ velocity: 1.2
20
+ time_anomaly: 0.6
21
+ graph_anomaly: 1.0
22
+ retry: 0.8
23
+ kyc: 0.5
24
+ user_risk: 0.8
25
+ random_seed: 2
26
+ difficulty: easy
27
+ release_mode: easy
28
+ seed: 2
29
+ fast_mode: false
30
+ n_checkpoints: 8
data/easy/seed_2/matched_pairs.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:583ee815999e1deade36980ead3d9f8cc93a23303e607b925728e153e2e250bd
3
+ size 84709
data/easy/seed_2/schema.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transactions_columns": {
3
+ "txn_id": "int32",
4
+ "sender_id": "int64",
5
+ "receiver_id": "int64",
6
+ "amount": "float32",
7
+ "timestamp": "float32",
8
+ "txn_type": "int8",
9
+ "is_fraud": "int8",
10
+ "fraud_type": "str",
11
+ "is_retry": "int8",
12
+ "risk_score": "float32",
13
+ "fail_prob": "float32",
14
+ "failed": "int8",
15
+ "twin_pair_id": "int64",
16
+ "template_id": "int64",
17
+ "twin_role": "str",
18
+ "twin_label": "int8",
19
+ "motif_source": "int8",
20
+ "motif_chain_state": "float32",
21
+ "motif_strength": "float32",
22
+ "dynamic_fraud_state": "float32",
23
+ "fraud_source": "str",
24
+ "motif_hit_count": "int32",
25
+ "trigger_event_idx": "int32",
26
+ "label_event_idx": "int32",
27
+ "label_delay": "int32",
28
+ "is_fallback_label": "int8",
29
+ "risk_noisy": "float32",
30
+ "neighbor_score": "float32",
31
+ "pair_freq": "float32",
32
+ "txn_count_10": "float32",
33
+ "amount_sum_10": "float32"
34
+ },
35
+ "matched_pairs_columns": {
36
+ "pair_event_id": "int64",
37
+ "twin_pair_id": "int64",
38
+ "template_id": "int64",
39
+ "matched_local_event_idx": "int64",
40
+ "prefix_txn_count": "int64",
41
+ "sender_id": "int64",
42
+ "label": "int64",
43
+ "twin_role": "str",
44
+ "matched_sender_id": "int64",
45
+ "total_txn_count": "int64",
46
+ "eval_timestamp": "float64",
47
+ "account_age": "float64",
48
+ "active_age": "float64",
49
+ "benchmark_mode": "str",
50
+ "difficulty": "str",
51
+ "seed": "int64"
52
+ },
53
+ "audit_summary_columns": {
54
+ "pair_total_txn_count_diff_mean": "float64",
55
+ "pair_total_txn_count_diff_max": "float64",
56
+ "auc_total_txn_count": "float64",
57
+ "auc_local_event_idx": "float64",
58
+ "auc_prefix_txn_count": "float64",
59
+ "auc_timestamp": "float64",
60
+ "auc_account_age": "float64",
61
+ "auc_active_age": "float64",
62
+ "fraud_label_event_idx_mean": "float64",
63
+ "fraud_label_event_idx_max": "float64",
64
+ "benign_eval_event_idx_mean": "float64",
65
+ "benign_eval_event_idx_max": "float64",
66
+ "pair_event_idx_diff_mean": "float64",
67
+ "pair_event_idx_diff_max": "float64",
68
+ "pair_active_age_diff_mean": "float64",
69
+ "pair_active_age_diff_max": "float64",
70
+ "pair_timestamp_diff_mean": "float64",
71
+ "pair_timestamp_diff_max": "float64",
72
+ "benign_motif_hit_rate": "float64",
73
+ "benign_motif_hit_pairs": "int64",
74
+ "matched_control_examples": "int64",
75
+ "matched_control_pair_events": "int64"
76
+ },
77
+ "files": [
78
+ "transactions.parquet",
79
+ "matched_pairs.parquet",
80
+ "audit_summary.csv",
81
+ "schema.json",
82
+ "config.yaml"
83
+ ]
84
+ }
data/easy/seed_2/transactions.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c61849d3d9679ef08fe1f595ced19145f2ecf60c15afc5c1d7998eb245b48f91
3
+ size 1289176
data/easy/seed_3/audit_summary.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ pair_total_txn_count_diff_mean,pair_total_txn_count_diff_max,auc_total_txn_count,auc_local_event_idx,auc_prefix_txn_count,auc_timestamp,auc_account_age,auc_active_age,fraud_label_event_idx_mean,fraud_label_event_idx_max,benign_eval_event_idx_mean,benign_eval_event_idx_max,pair_event_idx_diff_mean,pair_event_idx_diff_max,pair_active_age_diff_mean,pair_active_age_diff_max,pair_timestamp_diff_mean,pair_timestamp_diff_max,benign_motif_hit_rate,benign_motif_hit_pairs,matched_control_examples,matched_control_pair_events
2
+ 0.0,0.0,0.5,0.5,0.5,0.5,0.49999999999999994,0.49999999999999994,38.321280379371665,105.0,38.321280379371665,105.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0,3374,1687
data/easy/seed_3/config.yaml ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_users: 350
2
+ simulation_days: 45
3
+ fraud_ratio: 0.05
4
+ benchmark_mode: temporal_twins
5
+ user_params:
6
+ lambda_mean: 5.0
7
+ lambda_std: 1.0
8
+ mu_mean: 7.5
9
+ mu_std: 1.0
10
+ sigma_mean: 0.5
11
+ sigma_std: 0.2
12
+ upi_limits:
13
+ max_txn_amount: 100000.0
14
+ daily_limit: 100000.0
15
+ risk_model:
16
+ weights:
17
+ amount_ratio: 1.0
18
+ daily_ratio: 0.8
19
+ velocity: 1.2
20
+ time_anomaly: 0.6
21
+ graph_anomaly: 1.0
22
+ retry: 0.8
23
+ kyc: 0.5
24
+ user_risk: 0.8
25
+ random_seed: 3
26
+ difficulty: easy
27
+ release_mode: easy
28
+ seed: 3
29
+ fast_mode: false
30
+ n_checkpoints: 8
data/easy/seed_3/matched_pairs.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6844de17f7fb80b858563672a15dfb38102e0dc1f370a9690b7c0f2ae5a66145
3
+ size 81281
data/easy/seed_3/schema.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transactions_columns": {
3
+ "txn_id": "int32",
4
+ "sender_id": "int64",
5
+ "receiver_id": "int64",
6
+ "amount": "float32",
7
+ "timestamp": "float32",
8
+ "txn_type": "int8",
9
+ "is_fraud": "int8",
10
+ "fraud_type": "str",
11
+ "is_retry": "int8",
12
+ "risk_score": "float32",
13
+ "fail_prob": "float32",
14
+ "failed": "int8",
15
+ "twin_pair_id": "int64",
16
+ "template_id": "int64",
17
+ "twin_role": "str",
18
+ "twin_label": "int8",
19
+ "motif_source": "int8",
20
+ "motif_chain_state": "float32",
21
+ "motif_strength": "float32",
22
+ "dynamic_fraud_state": "float32",
23
+ "fraud_source": "str",
24
+ "motif_hit_count": "int32",
25
+ "trigger_event_idx": "int32",
26
+ "label_event_idx": "int32",
27
+ "label_delay": "int32",
28
+ "is_fallback_label": "int8",
29
+ "risk_noisy": "float32",
30
+ "neighbor_score": "float32",
31
+ "pair_freq": "float32",
32
+ "txn_count_10": "float32",
33
+ "amount_sum_10": "float32"
34
+ },
35
+ "matched_pairs_columns": {
36
+ "pair_event_id": "int64",
37
+ "twin_pair_id": "int64",
38
+ "template_id": "int64",
39
+ "matched_local_event_idx": "int64",
40
+ "prefix_txn_count": "int64",
41
+ "sender_id": "int64",
42
+ "label": "int64",
43
+ "twin_role": "str",
44
+ "matched_sender_id": "int64",
45
+ "total_txn_count": "int64",
46
+ "eval_timestamp": "float64",
47
+ "account_age": "float64",
48
+ "active_age": "float64",
49
+ "benchmark_mode": "str",
50
+ "difficulty": "str",
51
+ "seed": "int64"
52
+ },
53
+ "audit_summary_columns": {
54
+ "pair_total_txn_count_diff_mean": "float64",
55
+ "pair_total_txn_count_diff_max": "float64",
56
+ "auc_total_txn_count": "float64",
57
+ "auc_local_event_idx": "float64",
58
+ "auc_prefix_txn_count": "float64",
59
+ "auc_timestamp": "float64",
60
+ "auc_account_age": "float64",
61
+ "auc_active_age": "float64",
62
+ "fraud_label_event_idx_mean": "float64",
63
+ "fraud_label_event_idx_max": "float64",
64
+ "benign_eval_event_idx_mean": "float64",
65
+ "benign_eval_event_idx_max": "float64",
66
+ "pair_event_idx_diff_mean": "float64",
67
+ "pair_event_idx_diff_max": "float64",
68
+ "pair_active_age_diff_mean": "float64",
69
+ "pair_active_age_diff_max": "float64",
70
+ "pair_timestamp_diff_mean": "float64",
71
+ "pair_timestamp_diff_max": "float64",
72
+ "benign_motif_hit_rate": "float64",
73
+ "benign_motif_hit_pairs": "int64",
74
+ "matched_control_examples": "int64",
75
+ "matched_control_pair_events": "int64"
76
+ },
77
+ "files": [
78
+ "transactions.parquet",
79
+ "matched_pairs.parquet",
80
+ "audit_summary.csv",
81
+ "schema.json",
82
+ "config.yaml"
83
+ ]
84
+ }
data/easy/seed_3/transactions.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d4b28ff442286708fcc9fe214e53187c5a23d14f634b1387f4e2527c2e9ccd25
3
+ size 1357249
data/easy/seed_4/audit_summary.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ pair_total_txn_count_diff_mean,pair_total_txn_count_diff_max,auc_total_txn_count,auc_local_event_idx,auc_prefix_txn_count,auc_timestamp,auc_account_age,auc_active_age,fraud_label_event_idx_mean,fraud_label_event_idx_max,benign_eval_event_idx_mean,benign_eval_event_idx_max,pair_event_idx_diff_mean,pair_event_idx_diff_max,pair_active_age_diff_mean,pair_active_age_diff_max,pair_timestamp_diff_mean,pair_timestamp_diff_max,benign_motif_hit_rate,benign_motif_hit_pairs,matched_control_examples,matched_control_pair_events
2
+ 0.0,0.0,0.5,0.49999999999999994,0.49999999999999994,0.49999999999999994,0.5,0.5,39.30157766990291,100.0,39.30157766990291,100.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0,3296,1648
data/easy/seed_4/config.yaml ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_users: 350
2
+ simulation_days: 45
3
+ fraud_ratio: 0.05
4
+ benchmark_mode: temporal_twins
5
+ user_params:
6
+ lambda_mean: 5.0
7
+ lambda_std: 1.0
8
+ mu_mean: 7.5
9
+ mu_std: 1.0
10
+ sigma_mean: 0.5
11
+ sigma_std: 0.2
12
+ upi_limits:
13
+ max_txn_amount: 100000.0
14
+ daily_limit: 100000.0
15
+ risk_model:
16
+ weights:
17
+ amount_ratio: 1.0
18
+ daily_ratio: 0.8
19
+ velocity: 1.2
20
+ time_anomaly: 0.6
21
+ graph_anomaly: 1.0
22
+ retry: 0.8
23
+ kyc: 0.5
24
+ user_risk: 0.8
25
+ random_seed: 4
26
+ difficulty: easy
27
+ release_mode: easy
28
+ seed: 4
29
+ fast_mode: false
30
+ n_checkpoints: 8
data/easy/seed_4/matched_pairs.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:086f1c0336b7a670b6279207751fc337cf7d0a30cad24d4ae1b98542e94a4884
3
+ size 81766
data/easy/seed_4/schema.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transactions_columns": {
3
+ "txn_id": "int32",
4
+ "sender_id": "int64",
5
+ "receiver_id": "int64",
6
+ "amount": "float32",
7
+ "timestamp": "float64",
8
+ "txn_type": "int8",
9
+ "is_fraud": "int8",
10
+ "fraud_type": "str",
11
+ "is_retry": "int8",
12
+ "risk_score": "float32",
13
+ "fail_prob": "float32",
14
+ "failed": "int8",
15
+ "twin_pair_id": "int64",
16
+ "template_id": "int64",
17
+ "twin_role": "str",
18
+ "twin_label": "int8",
19
+ "motif_source": "int8",
20
+ "motif_chain_state": "float32",
21
+ "motif_strength": "float32",
22
+ "dynamic_fraud_state": "float32",
23
+ "fraud_source": "str",
24
+ "motif_hit_count": "int32",
25
+ "trigger_event_idx": "int32",
26
+ "label_event_idx": "int32",
27
+ "label_delay": "int32",
28
+ "is_fallback_label": "int8",
29
+ "risk_noisy": "float32",
30
+ "neighbor_score": "float32",
31
+ "pair_freq": "float32",
32
+ "txn_count_10": "float32",
33
+ "amount_sum_10": "float32"
34
+ },
35
+ "matched_pairs_columns": {
36
+ "pair_event_id": "int64",
37
+ "twin_pair_id": "int64",
38
+ "template_id": "int64",
39
+ "matched_local_event_idx": "int64",
40
+ "prefix_txn_count": "int64",
41
+ "sender_id": "int64",
42
+ "label": "int64",
43
+ "twin_role": "str",
44
+ "matched_sender_id": "int64",
45
+ "total_txn_count": "int64",
46
+ "eval_timestamp": "float64",
47
+ "account_age": "float64",
48
+ "active_age": "float64",
49
+ "benchmark_mode": "str",
50
+ "difficulty": "str",
51
+ "seed": "int64"
52
+ },
53
+ "audit_summary_columns": {
54
+ "pair_total_txn_count_diff_mean": "float64",
55
+ "pair_total_txn_count_diff_max": "float64",
56
+ "auc_total_txn_count": "float64",
57
+ "auc_local_event_idx": "float64",
58
+ "auc_prefix_txn_count": "float64",
59
+ "auc_timestamp": "float64",
60
+ "auc_account_age": "float64",
61
+ "auc_active_age": "float64",
62
+ "fraud_label_event_idx_mean": "float64",
63
+ "fraud_label_event_idx_max": "float64",
64
+ "benign_eval_event_idx_mean": "float64",
65
+ "benign_eval_event_idx_max": "float64",
66
+ "pair_event_idx_diff_mean": "float64",
67
+ "pair_event_idx_diff_max": "float64",
68
+ "pair_active_age_diff_mean": "float64",
69
+ "pair_active_age_diff_max": "float64",
70
+ "pair_timestamp_diff_mean": "float64",
71
+ "pair_timestamp_diff_max": "float64",
72
+ "benign_motif_hit_rate": "float64",
73
+ "benign_motif_hit_pairs": "int64",
74
+ "matched_control_examples": "int64",
75
+ "matched_control_pair_events": "int64"
76
+ },
77
+ "files": [
78
+ "transactions.parquet",
79
+ "matched_pairs.parquet",
80
+ "audit_summary.csv",
81
+ "schema.json",
82
+ "config.yaml"
83
+ ]
84
+ }
data/easy/seed_4/transactions.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f12641c7b1496ea8a84f9d26fccd9a65f9fdbd4678de0c821a472dfbb6ca9c58
3
+ size 1226851
data/hard/.DS_Store ADDED
Binary file (10.2 kB). View file
 
data/hard/seed_0/audit_summary.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ pair_total_txn_count_diff_mean,pair_total_txn_count_diff_max,auc_total_txn_count,auc_local_event_idx,auc_prefix_txn_count,auc_timestamp,auc_account_age,auc_active_age,fraud_label_event_idx_mean,fraud_label_event_idx_max,benign_eval_event_idx_mean,benign_eval_event_idx_max,pair_event_idx_diff_mean,pair_event_idx_diff_max,pair_active_age_diff_mean,pair_active_age_diff_max,pair_timestamp_diff_mean,pair_timestamp_diff_max,benign_motif_hit_rate,benign_motif_hit_pairs,matched_control_examples,matched_control_pair_events
2
+ 0.0,0.0,0.5,0.49999999999999994,0.49999999999999994,0.5000000000000001,0.5000000000000001,0.5000000000000001,72.17846153846153,835.0,72.17846153846153,835.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0,2600,1300
data/hard/seed_0/config.yaml ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_users: 350
2
+ simulation_days: 45
3
+ fraud_ratio: 0.05
4
+ benchmark_mode: temporal_twins
5
+ user_params:
6
+ lambda_mean: 5.0
7
+ lambda_std: 1.0
8
+ mu_mean: 7.5
9
+ mu_std: 1.0
10
+ sigma_mean: 0.5
11
+ sigma_std: 0.2
12
+ upi_limits:
13
+ max_txn_amount: 100000.0
14
+ daily_limit: 100000.0
15
+ risk_model:
16
+ weights:
17
+ amount_ratio: 1.0
18
+ daily_ratio: 0.8
19
+ velocity: 1.2
20
+ time_anomaly: 0.6
21
+ graph_anomaly: 1.0
22
+ retry: 0.8
23
+ kyc: 0.5
24
+ user_risk: 0.8
25
+ random_seed: 0
26
+ difficulty: hard
27
+ release_mode: hard
28
+ seed: 0
29
+ fast_mode: false
30
+ n_checkpoints: 8
data/hard/seed_0/matched_pairs.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cd7ab1965e4b8961bf5a314170a7ed2a33211b5006ce2ac3fe27327d1cfdaf5d
3
+ size 69658
data/hard/seed_0/schema.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transactions_columns": {
3
+ "txn_id": "int32",
4
+ "sender_id": "int64",
5
+ "receiver_id": "int64",
6
+ "amount": "float32",
7
+ "timestamp": "float32",
8
+ "txn_type": "int8",
9
+ "is_fraud": "int8",
10
+ "fraud_type": "str",
11
+ "is_retry": "int8",
12
+ "risk_score": "float32",
13
+ "fail_prob": "float32",
14
+ "failed": "int8",
15
+ "twin_pair_id": "int64",
16
+ "template_id": "int64",
17
+ "twin_role": "str",
18
+ "twin_label": "int8",
19
+ "motif_source": "int8",
20
+ "motif_chain_state": "float32",
21
+ "motif_strength": "float32",
22
+ "dynamic_fraud_state": "float32",
23
+ "fraud_source": "str",
24
+ "motif_hit_count": "int32",
25
+ "trigger_event_idx": "int32",
26
+ "label_event_idx": "int32",
27
+ "label_delay": "int32",
28
+ "is_fallback_label": "int8",
29
+ "risk_noisy": "float32",
30
+ "neighbor_score": "float32",
31
+ "pair_freq": "float32",
32
+ "txn_count_10": "float32",
33
+ "amount_sum_10": "float32"
34
+ },
35
+ "matched_pairs_columns": {
36
+ "pair_event_id": "int64",
37
+ "twin_pair_id": "int64",
38
+ "template_id": "int64",
39
+ "matched_local_event_idx": "int64",
40
+ "prefix_txn_count": "int64",
41
+ "sender_id": "int64",
42
+ "label": "int64",
43
+ "twin_role": "str",
44
+ "matched_sender_id": "int64",
45
+ "total_txn_count": "int64",
46
+ "eval_timestamp": "float64",
47
+ "account_age": "float64",
48
+ "active_age": "float64",
49
+ "benchmark_mode": "str",
50
+ "difficulty": "str",
51
+ "seed": "int64"
52
+ },
53
+ "audit_summary_columns": {
54
+ "pair_total_txn_count_diff_mean": "float64",
55
+ "pair_total_txn_count_diff_max": "float64",
56
+ "auc_total_txn_count": "float64",
57
+ "auc_local_event_idx": "float64",
58
+ "auc_prefix_txn_count": "float64",
59
+ "auc_timestamp": "float64",
60
+ "auc_account_age": "float64",
61
+ "auc_active_age": "float64",
62
+ "fraud_label_event_idx_mean": "float64",
63
+ "fraud_label_event_idx_max": "float64",
64
+ "benign_eval_event_idx_mean": "float64",
65
+ "benign_eval_event_idx_max": "float64",
66
+ "pair_event_idx_diff_mean": "float64",
67
+ "pair_event_idx_diff_max": "float64",
68
+ "pair_active_age_diff_mean": "float64",
69
+ "pair_active_age_diff_max": "float64",
70
+ "pair_timestamp_diff_mean": "float64",
71
+ "pair_timestamp_diff_max": "float64",
72
+ "benign_motif_hit_rate": "float64",
73
+ "benign_motif_hit_pairs": "int64",
74
+ "matched_control_examples": "int64",
75
+ "matched_control_pair_events": "int64"
76
+ },
77
+ "files": [
78
+ "transactions.parquet",
79
+ "matched_pairs.parquet",
80
+ "audit_summary.csv",
81
+ "schema.json",
82
+ "config.yaml"
83
+ ]
84
+ }
data/hard/seed_0/transactions.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5c6087b6fbbf6fe37ebf59e9e8c364c54ec1ebe6b6d7bd26add4ebcd819f68f0
3
+ size 2281421
data/hard/seed_1/audit_summary.csv ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ pair_total_txn_count_diff_mean,pair_total_txn_count_diff_max,auc_total_txn_count,auc_local_event_idx,auc_prefix_txn_count,auc_timestamp,auc_account_age,auc_active_age,fraud_label_event_idx_mean,fraud_label_event_idx_max,benign_eval_event_idx_mean,benign_eval_event_idx_max,pair_event_idx_diff_mean,pair_event_idx_diff_max,pair_active_age_diff_mean,pair_active_age_diff_max,pair_timestamp_diff_mean,pair_timestamp_diff_max,benign_motif_hit_rate,benign_motif_hit_pairs,matched_control_examples,matched_control_pair_events
2
+ 0.0,0.0,0.49999999999999994,0.5,0.5,0.49999999999999994,0.49999999999999994,0.49999999999999994,92.03990963855422,1822.0,92.03990963855422,1822.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0,2656,1328
data/hard/seed_1/config.yaml ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ num_users: 350
2
+ simulation_days: 45
3
+ fraud_ratio: 0.05
4
+ benchmark_mode: temporal_twins
5
+ user_params:
6
+ lambda_mean: 5.0
7
+ lambda_std: 1.0
8
+ mu_mean: 7.5
9
+ mu_std: 1.0
10
+ sigma_mean: 0.5
11
+ sigma_std: 0.2
12
+ upi_limits:
13
+ max_txn_amount: 100000.0
14
+ daily_limit: 100000.0
15
+ risk_model:
16
+ weights:
17
+ amount_ratio: 1.0
18
+ daily_ratio: 0.8
19
+ velocity: 1.2
20
+ time_anomaly: 0.6
21
+ graph_anomaly: 1.0
22
+ retry: 0.8
23
+ kyc: 0.5
24
+ user_risk: 0.8
25
+ random_seed: 1
26
+ difficulty: hard
27
+ release_mode: hard
28
+ seed: 1
29
+ fast_mode: false
30
+ n_checkpoints: 8
data/hard/seed_1/matched_pairs.parquet ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:69211d9c6d5c6d43b1b2e3ce897c152b2928bd7ff579032af63b398661888fa7
3
+ size 70787
data/hard/seed_1/schema.json ADDED
@@ -0,0 +1,84 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transactions_columns": {
3
+ "txn_id": "int32",
4
+ "sender_id": "int64",
5
+ "receiver_id": "int64",
6
+ "amount": "float32",
7
+ "timestamp": "float32",
8
+ "txn_type": "int8",
9
+ "is_fraud": "int8",
10
+ "fraud_type": "str",
11
+ "is_retry": "int8",
12
+ "risk_score": "float32",
13
+ "fail_prob": "float32",
14
+ "failed": "int8",
15
+ "twin_pair_id": "int64",
16
+ "template_id": "int64",
17
+ "twin_role": "str",
18
+ "twin_label": "int8",
19
+ "motif_source": "int8",
20
+ "motif_chain_state": "float32",
21
+ "motif_strength": "float32",
22
+ "dynamic_fraud_state": "float32",
23
+ "fraud_source": "str",
24
+ "motif_hit_count": "int32",
25
+ "trigger_event_idx": "int32",
26
+ "label_event_idx": "int32",
27
+ "label_delay": "int32",
28
+ "is_fallback_label": "int8",
29
+ "risk_noisy": "float32",
30
+ "neighbor_score": "float32",
31
+ "pair_freq": "float32",
32
+ "txn_count_10": "float32",
33
+ "amount_sum_10": "float32"
34
+ },
35
+ "matched_pairs_columns": {
36
+ "pair_event_id": "int64",
37
+ "twin_pair_id": "int64",
38
+ "template_id": "int64",
39
+ "matched_local_event_idx": "int64",
40
+ "prefix_txn_count": "int64",
41
+ "sender_id": "int64",
42
+ "label": "int64",
43
+ "twin_role": "str",
44
+ "matched_sender_id": "int64",
45
+ "total_txn_count": "int64",
46
+ "eval_timestamp": "float64",
47
+ "account_age": "float64",
48
+ "active_age": "float64",
49
+ "benchmark_mode": "str",
50
+ "difficulty": "str",
51
+ "seed": "int64"
52
+ },
53
+ "audit_summary_columns": {
54
+ "pair_total_txn_count_diff_mean": "float64",
55
+ "pair_total_txn_count_diff_max": "float64",
56
+ "auc_total_txn_count": "float64",
57
+ "auc_local_event_idx": "float64",
58
+ "auc_prefix_txn_count": "float64",
59
+ "auc_timestamp": "float64",
60
+ "auc_account_age": "float64",
61
+ "auc_active_age": "float64",
62
+ "fraud_label_event_idx_mean": "float64",
63
+ "fraud_label_event_idx_max": "float64",
64
+ "benign_eval_event_idx_mean": "float64",
65
+ "benign_eval_event_idx_max": "float64",
66
+ "pair_event_idx_diff_mean": "float64",
67
+ "pair_event_idx_diff_max": "float64",
68
+ "pair_active_age_diff_mean": "float64",
69
+ "pair_active_age_diff_max": "float64",
70
+ "pair_timestamp_diff_mean": "float64",
71
+ "pair_timestamp_diff_max": "float64",
72
+ "benign_motif_hit_rate": "float64",
73
+ "benign_motif_hit_pairs": "int64",
74
+ "matched_control_examples": "int64",
75
+ "matched_control_pair_events": "int64"
76
+ },
77
+ "files": [
78
+ "transactions.parquet",
79
+ "matched_pairs.parquet",
80
+ "audit_summary.csv",
81
+ "schema.json",
82
+ "config.yaml"
83
+ ]
84
+ }