Datasets:

samreed-research
/

billsim

@@ -39,7 +39,10 @@
   },
   "@type": "sc:Dataset",
   "name": "BillSim",
-  "conformsTo": "http://mlcommons.org/croissant/1.0",
   "description": "BillSim is a synthetic heterogeneous graph benchmark for non-adversarial anomaly detection in usage-based billing pipelines. The default dataset contains 63,155 nodes across 11 entity types (customer, contract, sku, pricing_rule, usage_record, rated_charge, invoice, line_item, payment, ar_record, gl_entry) and 108,930 edges across 14 relationship types (subscribes_to, has_sku, governs, applies_to, generates_usage, for_sku, rates_to, priced_by, billed_on, part_of, invoiced_to, settles, opens_ar, posts_to). Six anomaly types are organized into three categories: structural (Pricing Mismatch, Revenue Recognition, Premature AR Closure), relational (Missing Metering, Orphaned Invoice Lines), and peer-comparison (Stale Discount). Five baselines are evaluated (MLP, GCN, GAT, RGCN, HGT) across 10 random seeds and 4 injection rates (0.01, 0.03, 0.05, 0.10). Headline finding: a 0.374 AUROC oracle gap on peer-comparison anomalies between a hand-coded oracle (0.965 AUROC on Stale Discount) and the best message-passing baseline (RGCN at 0.591; HGT at 0.540 is statistically indistinguishable), localizing the limitation to message-passing aggregation rather than task ill-posedness. Distributed as a PyTorch Geometric HeteroData (.pt) tensor plus deterministic Python generation scripts.",
   "citeAs": "Anonymous Authors. BillSim: A Synthetic Heterogeneous Graph Benchmark for Non-Adversarial Anomaly Detection in Usage-Based Billing. Submitted to NeurIPS 2026 Evaluations & Datasets Track, 2026.",
   "license": "https://creativecommons.org/licenses/by/4.0/",
@@ -59,7 +62,7 @@
     "email": "anon@example.com"
   },
   "rai:dataCollection": "Fully synthetic dataset generated by a deterministic Python pipeline. No real customer, account, transaction, or other financial data was collected, sampled, scraped, or otherwise sourced from any real-world subject or system. Every node, edge, feature, and label is produced by parameterized statistical distributions (Pareto for revenue, log-normal for usage volumes, Poisson for arrivals, uniform for pricing) seeded with a NumPy random state for bit-level reproducibility.",
-  "rai:dataCollectionType": "Synthetic data generation",
   "rai:dataPreprocessingProtocol": "Per-node-type z-score normalization (mean 0, std 1) on each numeric feature column. Categorical features (e.g., customer segment, payment terms, SKU category, payment method) are integer-encoded with stable per-process hashing. Train/test split is deterministic per seed: static entity types (customer, contract, pricing_rule, sku) use a 75/25 random split; dynamic entity types (usage_record, rated_charge, invoice, line_item, payment, ar_record, gl_entry) use a temporal split of months 1-18 (train) versus 19-24 (test) over a 24-month simulated horizon. No missing values, no NaN, no Inf, and no out-of-bounds edge indices are present; integrity is verified by an automated check on every regeneration.",
   "rai:dataAnnotationProtocol": "Anomaly labels are generated programmatically at injection time, not annotated by humans. The injector marks each affected node with a binary anomaly indicator (y in {0, 1}) and an integer anomaly type (anomaly_type in {0, 1, 2, 3, 4, 5, 6}). Every node has at most one anomaly type assigned, so the multi-class label set is mutually exclusive across the six anomaly categories.",
   "rai:dataUseCases": "Intended use: benchmarking graph neural network architectures on non-adversarial anomaly detection in heterogeneous temporal graphs, and evaluating which graph-reasoning patterns (structural degree counting, multi-hop path matching, relation-specific traversal, peer comparison) different architectures excel at. Not intended for fraud detection: anomalies in BillSim are operational and configuration errors, not adversarial attack patterns; fraud involves intentional concealment, billing errors do not. Not intended for production billing system deployment: the synthetic distributions deliberately do not replicate real billing data complexity.",

   },
   "@type": "sc:Dataset",
   "name": "BillSim",
+  "conformsTo": [
+    "http://mlcommons.org/croissant/1.0",
+    "http://mlcommons.org/croissant/RAI/1.0"
+  ],
   "description": "BillSim is a synthetic heterogeneous graph benchmark for non-adversarial anomaly detection in usage-based billing pipelines. The default dataset contains 63,155 nodes across 11 entity types (customer, contract, sku, pricing_rule, usage_record, rated_charge, invoice, line_item, payment, ar_record, gl_entry) and 108,930 edges across 14 relationship types (subscribes_to, has_sku, governs, applies_to, generates_usage, for_sku, rates_to, priced_by, billed_on, part_of, invoiced_to, settles, opens_ar, posts_to). Six anomaly types are organized into three categories: structural (Pricing Mismatch, Revenue Recognition, Premature AR Closure), relational (Missing Metering, Orphaned Invoice Lines), and peer-comparison (Stale Discount). Five baselines are evaluated (MLP, GCN, GAT, RGCN, HGT) across 10 random seeds and 4 injection rates (0.01, 0.03, 0.05, 0.10). Headline finding: a 0.374 AUROC oracle gap on peer-comparison anomalies between a hand-coded oracle (0.965 AUROC on Stale Discount) and the best message-passing baseline (RGCN at 0.591; HGT at 0.540 is statistically indistinguishable), localizing the limitation to message-passing aggregation rather than task ill-posedness. Distributed as a PyTorch Geometric HeteroData (.pt) tensor plus deterministic Python generation scripts.",
   "citeAs": "Anonymous Authors. BillSim: A Synthetic Heterogeneous Graph Benchmark for Non-Adversarial Anomaly Detection in Usage-Based Billing. Submitted to NeurIPS 2026 Evaluations & Datasets Track, 2026.",
   "license": "https://creativecommons.org/licenses/by/4.0/",
     "email": "anon@example.com"
   },
   "rai:dataCollection": "Fully synthetic dataset generated by a deterministic Python pipeline. No real customer, account, transaction, or other financial data was collected, sampled, scraped, or otherwise sourced from any real-world subject or system. Every node, edge, feature, and label is produced by parameterized statistical distributions (Pareto for revenue, log-normal for usage volumes, Poisson for arrivals, uniform for pricing) seeded with a NumPy random state for bit-level reproducibility.",
+  "rai:dataCollectionType": "Software Collection",
   "rai:dataPreprocessingProtocol": "Per-node-type z-score normalization (mean 0, std 1) on each numeric feature column. Categorical features (e.g., customer segment, payment terms, SKU category, payment method) are integer-encoded with stable per-process hashing. Train/test split is deterministic per seed: static entity types (customer, contract, pricing_rule, sku) use a 75/25 random split; dynamic entity types (usage_record, rated_charge, invoice, line_item, payment, ar_record, gl_entry) use a temporal split of months 1-18 (train) versus 19-24 (test) over a 24-month simulated horizon. No missing values, no NaN, no Inf, and no out-of-bounds edge indices are present; integrity is verified by an automated check on every regeneration.",
   "rai:dataAnnotationProtocol": "Anomaly labels are generated programmatically at injection time, not annotated by humans. The injector marks each affected node with a binary anomaly indicator (y in {0, 1}) and an integer anomaly type (anomaly_type in {0, 1, 2, 3, 4, 5, 6}). Every node has at most one anomaly type assigned, so the multi-class label set is mutually exclusive across the six anomaly categories.",
   "rai:dataUseCases": "Intended use: benchmarking graph neural network architectures on non-adversarial anomaly detection in heterogeneous temporal graphs, and evaluating which graph-reasoning patterns (structural degree counting, multi-hop path matching, relation-specific traversal, peer comparison) different architectures excel at. Not intended for fraud detection: anomalies in BillSim are operational and configuration errors, not adversarial attack patterns; fraud involves intentional concealment, billing errors do not. Not intended for production billing system deployment: the synthetic distributions deliberately do not replicate real billing data complexity.",