| """ |
| Generate a tight, comprehensive IEEE-style PDF paper using fpdf2. |
| Target: 12 pages. No wasted whitespace. Redundant figures removed. |
| """ |
| import os, sys |
| sys.path.insert(0, '/app/fraud_detection') |
| from fpdf import FPDF |
|
|
| FIGURES_DIR = '/app/fraud_detection/figures' |
| PAPER_DIR = '/app/fraud_detection/paper' |
| os.makedirs(PAPER_DIR, exist_ok=True) |
|
|
| LM = 14 |
| RM = 14 |
| TM = 14 |
| BW = 215.9 - LM - RM |
|
|
|
|
| class IEEEPaper(FPDF): |
| def __init__(self): |
| super().__init__('P', 'mm', 'letter') |
| self.set_margins(LM, TM, RM) |
| self.set_auto_page_break(auto=True, margin=16) |
|
|
| def header(self): |
| if self.page_no() > 1: |
| self.set_font('Helvetica', 'I', 6.5) |
| self.cell(0, 3, 'IEEE -- Fraud Detection with Explainable AI', align='C') |
| self.ln(4) |
|
|
| def footer(self): |
| self.set_y(-12) |
| self.set_font('Helvetica', 'I', 7) |
| self.cell(0, 8, f'{self.page_no()}', align='C') |
|
|
| def sec(self, num, title): |
| self.ln(3) |
| self.set_font('Helvetica', 'B', 10) |
| t = f'{num}. {title.upper()}' if num else title.upper() |
| self.cell(0, 5, t, ln=True) |
| self.ln(1) |
|
|
| def sub(self, label, title): |
| self.ln(1.5) |
| self.set_font('Helvetica', 'B', 9) |
| self.cell(0, 4.5, f'{label} {title}', ln=True) |
| self.ln(0.5) |
|
|
| def p(self, text): |
| self.set_font('Times', '', 9) |
| self.multi_cell(0, 3.8, text) |
| self.ln(1) |
|
|
| def bullet(self, items): |
| self.set_font('Times', '', 9) |
| for item in items: |
| self.set_x(LM + 3) |
| self.cell(3, 3.8, '-') |
| self.multi_cell(BW - 6, 3.8, item) |
| self.ln(0.3) |
| self.ln(0.5) |
|
|
| def fig(self, path, caption, w=145): |
| if not os.path.exists(path): |
| return |
| self.ln(1.5) |
| x = (self.w - w) / 2 |
| self.image(path, x=x, w=w) |
| self.ln(1) |
| self.set_font('Helvetica', 'I', 7.5) |
| self.multi_cell(0, 3.5, caption, align='C') |
| self.ln(1.5) |
|
|
| def tbl(self, hdrs, rows, caption=''): |
| if caption: |
| self.ln(1) |
| self.set_font('Helvetica', 'I', 7.5) |
| self.multi_cell(0, 3.5, caption, align='C') |
| self.ln(0.5) |
| cw = BW / len(hdrs) |
| self.set_font('Helvetica', 'B', 7) |
| for h in hdrs: |
| self.cell(cw, 4, h, border=1, align='C') |
| self.ln() |
| self.set_font('Times', '', 7) |
| for row in rows: |
| for c in row: |
| self.cell(cw, 4, str(c), border=1, align='C') |
| self.ln() |
| self.ln(1.5) |
|
|
|
|
| def build(): |
| pdf = IEEEPaper() |
| F = FIGURES_DIR |
|
|
| |
| pdf.add_page() |
| pdf.ln(6) |
| pdf.set_font('Helvetica', 'B', 15) |
| pdf.multi_cell(0, 7.5, 'A Comprehensive Ensemble-Based Framework\nfor Credit Card Fraud Detection with Explainable AI', align='C') |
| pdf.ln(3) |
| pdf.set_font('Helvetica', '', 10) |
| pdf.cell(0, 5, 'Raj Vivan', align='C', ln=True) |
| pdf.set_font('Helvetica', 'I', 8.5) |
| pdf.cell(0, 4, 'Department of Computer Science | Independent Research', align='C', ln=True) |
| pdf.ln(4) |
|
|
| pdf.set_font('Helvetica', 'B', 9) |
| pdf.cell(0, 4, 'Abstract', align='C', ln=True) |
| pdf.ln(1) |
| pdf.p( |
| 'Credit card fraud poses a significant and growing threat to the global financial ecosystem, with estimated annual losses exceeding ' |
| '$32 billion. This paper presents a comprehensive, end-to-end fraud detection framework that systematically develops, evaluates, and ' |
| 'compares seven machine learning approaches: Logistic Regression, Random Forest, XGBoost, LightGBM, Multilayer Perceptron, ' |
| 'Autoencoder-based anomaly detection, and a Voting Ensemble. Using the European Cardholder benchmark dataset (284,807 transactions, ' |
| '0.173% fraud rate), we engineer 12 novel features and address class imbalance through SMOTE oversampling (applied exclusively after ' |
| 'train-test splitting) and cost-sensitive learning. XGBoost achieves the best performance with PR-AUC of 0.8166, precision of 0.9048, ' |
| 'recall of 0.8028, and F1 of 0.8507. Threshold optimization from 0.5 to 0.55 improves F1 to 0.8636. SHAP and LIME explainability ' |
| 'analysis identifies V4, V14, and PCA_magnitude as primary fraud discriminators. Error analysis reveals that false negatives arise ' |
| 'from sophisticated fraud closely mimicking legitimate behavior. The model is deployed as a FastAPI service with sub-10ms latency. ' |
| 'All code, models, and results are publicly available.' |
| ) |
| pdf.set_font('Helvetica', 'I', 7.5) |
| pdf.cell(0, 4, 'Keywords: fraud detection, XGBoost, ensemble learning, SHAP, LIME, class imbalance, SMOTE, anomaly detection', ln=True) |
|
|
| |
| pdf.sec('I', 'Introduction') |
| pdf.p( |
| 'Financial fraud detection has emerged as one of the most consequential applications of machine learning. The global shift toward ' |
| 'electronic payments has created unprecedented transaction volumes while simultaneously enabling sophisticated fraud. According to ' |
| 'the Nilson Report [21], worldwide card fraud losses reached $32.34 billion in 2021 (a 14% increase year-over-year), projected to ' |
| 'exceed $43 billion by 2026. The fundamental challenge lies in extreme class imbalance: fraudulent transactions typically constitute ' |
| 'less than 0.5% of all transactions, rendering accuracy metrics meaningless and necessitating PR-AUC, F1, and MCC [18].' |
| ) |
| pdf.p( |
| 'A second challenge is concept drift [1]: fraudsters continuously adapt, causing model performance to degrade over time. Previous ' |
| 'approaches range from rule-based expert systems [12] to deep learning architectures [13]. However, extensive benchmarking by ' |
| 'Shwartz-Ziv and Armon [9] and Grinsztajn et al. [10] demonstrates that well-tuned gradient-boosted trees consistently outperform ' |
| 'deep learning on tabular data, including fraud detection, when combined with thoughtful feature engineering.' |
| ) |
| pdf.p('This paper makes the following contributions:') |
| pdf.bullet([ |
| 'Systematic comparison of seven ML approaches spanning linear models, tree ensembles, neural networks, and anomaly detection.', |
| 'Novel feature engineering producing 12 features capturing temporal cycles, transaction velocity, and PCA interactions.', |
| 'Rigorous methodology: SMOTE only after splitting; scaler fitted on train only; six metrics including PR-AUC and MCC.', |
| 'SHAP (global) and LIME (local) explainability analysis identifying key fraud indicators.', |
| 'Production FastAPI deployment achieving sub-10ms latency with business impact quantification.', |
| ]) |
|
|
| |
| pdf.sec('II', 'Related Work') |
| pdf.p( |
| 'Bolton and Hand [12] provided an early survey of statistical fraud detection. Dal Pozzolo et al. [1] analyzed class imbalance and ' |
| 'concept drift in real-world systems, while their follow-up [22] investigated when undersampling is effective. Chawla et al. [2] ' |
| 'introduced SMOTE for synthetic minority oversampling; Fernandez et al. [3] later demonstrated that SMOTE must be applied exclusively ' |
| 'to training data to avoid data leakage that produces over-optimistic estimates.' |
| ) |
| pdf.p( |
| 'Tree-based methods dominate tabular fraud detection. Xuan et al. [17] showed Random Forests achieve robust baseline performance. ' |
| 'Chen and Guestrin [4] introduced XGBoost; Ke et al. [5] proposed LightGBM with leaf-wise growth and GOSS; Prokhorenkova et al. [16] ' |
| 'introduced CatBoost with ordered boosting. Taha and Malebary [14] demonstrated optimized LightGBM for fraud detection. ' |
| 'Pumsirirat and Yan [6] employed autoencoders trained on legitimate transactions only, detecting fraud via reconstruction error. ' |
| 'Zhang et al. [13] proposed attention-based RNNs for sequential patterns.' |
| ) |
| pdf.p( |
| 'For explainability, Lundberg and Lee [7] introduced SHAP based on Shapley values from cooperative game theory. Ribeiro et al. [8] ' |
| 'proposed LIME for instance-level interpretation via local linear approximations. Belle and Papantonis [15] surveyed XAI methods for ' |
| 'financial decision-making. Akiba et al. [11] introduced Optuna with TPE sampling for efficient hyperparameter optimization.' |
| ) |
|
|
| |
| pdf.sec('III', 'Dataset and Exploratory Data Analysis') |
| pdf.sub('A.', 'Dataset Description') |
| pdf.p( |
| 'We use the European Cardholder dataset [1], containing 284,807 transactions over two days in September 2013. Each transaction has ' |
| '28 PCA-transformed features (V1-V28), raw Time and Amount, and a binary Class label (0=legitimate, 1=fraud). The PCA transformation ' |
| 'protects cardholder privacy but prevents domain-specific feature engineering. The dataset has extreme class imbalance: only 492 ' |
| 'fraudulent transactions (0.173%), yielding an imbalance ratio of 1:577.' |
| ) |
| pdf.tbl( |
| ['Class', 'Count', 'Percentage', 'Ratio'], |
| [['Legitimate (0)', '284,315', '99.827%', '---'], |
| ['Fraud (1)', '492', '0.173%', '1:577'], |
| ['Total', '284,807', '100%', '---']], |
| 'Table I: Class Distribution' |
| ) |
| pdf.fig(os.path.join(F, 'class_distribution.png'), |
| 'Fig. 1. Class distribution showing extreme imbalance (0.173% fraud).', w=130) |
|
|
| pdf.sub('B.', 'Transaction Amount and Temporal Patterns') |
| pdf.p( |
| 'Legitimate transactions have a mean of $88.29 (median $22.00); fraudulent transactions have a higher mean of $122.21 but lower ' |
| 'median of $9.25. This bimodal fraud pattern suggests two strategies: (i) low-value "testing" transactions verifying stolen cards, ' |
| 'and (ii) moderate-to-high-value theft. The nighttime (0-6h) fraud rate is 0.518%, nearly 4x the daytime rate of 0.137%, consistent ' |
| 'with fraudsters exploiting low-monitoring periods. These patterns motivate our cyclic temporal and amount-based feature engineering.' |
| ) |
| pdf.fig(os.path.join(F, 'amount_analysis.png'), |
| 'Fig. 2. Amount analysis: (a) legitimate, (b) fraud, (c) log-scaled comparison, (d) boxplot.', w=145) |
| pdf.fig(os.path.join(F, 'time_analysis.png'), |
| 'Fig. 3. Temporal patterns: (a) transaction density by hour, (b) fraud rate by hour.', w=135) |
|
|
| pdf.sub('C.', 'Feature Correlations and Key Observations') |
| pdf.p( |
| 'Pearson correlation identifies V17 (r=-0.326), V14 (r=-0.303), and V12 (r=-0.261) as having the strongest negative correlation ' |
| 'with fraud; V11 (+0.155) and V4 (+0.133) show the strongest positive correlation. Amount has near-zero correlation (r=0.006), ' |
| 'confirming that amount-based rules alone would be ineffective. Five key observations: (1) the 1:577 imbalance makes accuracy ' |
| 'meaningless; (2) bimodal fraud amounts require engineered deviation features; (3) the 4x nighttime fraud rate provides temporal ' |
| 'signal; (4) V14, V17, V12, V4, V11 carry the strongest fraud signal; (5) no missing values exist, with 1,081 duplicates removed.' |
| ) |
| pdf.fig(os.path.join(F, 'correlation_heatmap.png'), |
| 'Fig. 4. Feature correlation with fraud class. Red bars indicate negative correlation (lower values signal fraud).', w=120) |
|
|
| |
| pdf.sec('IV', 'Methodology') |
| pdf.sub('A.', 'Feature Engineering') |
| pdf.p( |
| 'We augment the original 30 features with 12 engineered features (42 total). Temporal features: cyclic hour encoding ' |
| '(Hour_sin, Hour_cos) and Time_diff (inter-arrival time). Amount features: Amount_log, Amount_deviation_mean, ' |
| 'Amount_deviation_median, Amount_zscore, and Transaction_velocity. Interaction features: V14*V17, V12*V14, V10*V14 (capturing ' |
| 'joint effects between top discriminators). PCA_magnitude: L2 norm of all V features, measuring overall transaction abnormality.' |
| ) |
| pdf.sub('B.', 'Class Imbalance, Splitting, and Scaling') |
| pdf.p( |
| 'Two approaches are compared, both applied exclusively to training data. SMOTE [2] generates synthetic fraud at a 1:2 ratio ' |
| '(99,138 synthetic + 198,277 legitimate), used only for MLP training. Cost-sensitive learning applies balanced class weights ' |
| '(w0=0.501, w1=300.01) for tree models and Logistic Regression. Stratified 70/15/15 splitting preserves the fraud ratio across ' |
| 'all sets (Train: 198,608 samples/331 fraud; Val/Test: 42,559/71 each). RobustScaler normalizes by IQR, fitted on train only.' |
| ) |
| pdf.sub('C.', 'Model Descriptions') |
| pdf.p( |
| '(1) Logistic Regression: L2-regularized (C=0.1), balanced weights, interpretable baseline. ' |
| '(2) Random Forest: 150 trees, depth 12, balanced weights. ' |
| '(3) XGBoost: 200 estimators, depth 6, lr=0.1, scale_pos_weight from class frequencies, histogram splitting. ' |
| '(4) LightGBM: 200 estimators, depth 8, lr=0.05, leaf-wise growth with GOSS. ' |
| '(5) MLP: 128-64-32 neurons, ReLU, adaptive lr, early stopping, trained on SMOTE data. ' |
| '(6) Autoencoder: 42-64-32-16-32-64-42, trained 50 epochs on legitimate only, detects fraud via reconstruction error. ' |
| '(7) Voting Ensemble: soft voting over three best tuned models (XGBoost, LightGBM, RF).' |
| ) |
| pdf.sub('D.', 'Hyperparameter Optimization') |
| pdf.p( |
| 'Optuna [11] with TPE sampler tunes XGBoost, LightGBM, and Random Forest (15-20 trials each), optimizing PR-AUC on the validation set.' |
| ) |
| pdf.fig(os.path.join(F, 'architecture_diagram.png'), |
| 'Fig. 5. End-to-end system architecture from transaction input through inference to API output and monitoring.', w=145) |
|
|
| |
| pdf.sec('V', 'Experimental Setup') |
| pdf.p( |
| 'Experiments use Python 3.12 with scikit-learn 1.8.0, XGBoost 3.2.0, LightGBM 4.6.0, PyTorch 2.11.0, Optuna 4.8.0, SHAP 0.51.0, ' |
| 'and LIME 0.2.0.1 on CPU infrastructure (~25 min total training). Six metrics reported: Precision, Recall, F1, ROC-AUC, ' |
| 'PR-AUC (primary, most informative under extreme imbalance [18]), and MCC (balanced measure across all confusion matrix quadrants).' |
| ) |
|
|
| |
| pdf.sec('VI', 'Results and Discussion') |
| pdf.sub('A.', 'Model Comparison') |
| pdf.p('Table II presents comprehensive test set evaluation at threshold 0.5.') |
| pdf.tbl( |
| ['Model', 'Prec.', 'Recall', 'F1', 'ROC-AUC', 'PR-AUC', 'MCC'], |
| [ |
| ['XGBoost', '0.905', '0.803', '0.851', '0.974', '0.817', '0.852'], |
| ['Voting Ens.', '0.864', '0.803', '0.832', '0.978', '0.801', '0.832'], |
| ['LGBM Tuned', '0.707', '0.817', '0.758', '0.932', '0.796', '0.760'], |
| ['XGB Tuned', '0.838', '0.803', '0.820', '0.970', '0.793', '0.820'], |
| ['RF Tuned', '0.873', '0.775', '0.821', '0.968', '0.793', '0.822'], |
| ['Random Forest', '0.833', '0.775', '0.803', '0.953', '0.771', '0.803'], |
| ['MLP', '0.691', '0.789', '0.737', '0.943', '0.752', '0.738'], |
| ['Logistic Reg.', '0.049', '0.887', '0.092', '0.962', '0.735', '0.204'], |
| ['Autoencoder', '0.003', '1.000', '0.007', '0.960', '0.044', '0.041'], |
| ], |
| 'Table II: Comprehensive Model Comparison on Test Set (threshold=0.5)' |
| ) |
| pdf.p( |
| 'Observation 1 -- Tree models dominate: XGBoost achieves the highest PR-AUC (0.817), F1 (0.851), and MCC (0.852), confirming [9] ' |
| 'that gradient-boosted trees remain strongest for tabular data. The Voting Ensemble achieves marginally higher ROC-AUC (0.978) ' |
| 'but does not improve PR-AUC, suggesting insufficient member diversity.' |
| ) |
| pdf.p( |
| 'Observation 2 -- Precision-recall tradeoff: Logistic Regression achieves high recall (0.887) but catastrophic precision (0.049), ' |
| 'flagging 1,229 legitimate transactions. The aggressive class weight (300x) creates a boundary that is far too liberal, producing ' |
| 'a flood of false alarms that would overwhelm any operational team.' |
| ) |
| pdf.p( |
| 'Observation 3 -- Autoencoder failure: Perfect recall (1.0) but precision of only 0.003 (21,209 false positives). The PCA-transformed ' |
| 'space causes the autoencoder to reconstruct dominant variance directions; fraud signals in minor components produce similar ' |
| 'reconstruction errors to legitimate noise, making discrimination unreliable. PR-AUC of 0.044 is near-random.' |
| ) |
| pdf.p( |
| 'Observation 4 -- Tuning: Optuna dramatically improved LightGBM (PR-AUC 0.012 to 0.796 by correcting over-aggressive ' |
| 'scale_pos_weight). However, tuned XGBoost (0.793) slightly underperformed the base (0.817), suggesting the base was near-optimal.' |
| ) |
| pdf.fig(os.path.join(F, 'roc_curves.png'), |
| 'Fig. 6. ROC curves for top 5 models (all ROC-AUC > 0.93).', w=115) |
| pdf.fig(os.path.join(F, 'pr_curves.png'), |
| 'Fig. 7. Precision-Recall curves. XGBoost achieves the highest AP (0.817).', w=115) |
| pdf.fig(os.path.join(F, 'confusion_matrices.png'), |
| 'Fig. 8. Confusion matrices. XGBoost: 57 TP, 6 FP -- best balance.', w=150) |
|
|
| pdf.sub('B.', 'Threshold Optimization') |
| pdf.p( |
| 'A systematic threshold sweep (0.05 to 0.95) on XGBoost reveals an optimal threshold of 0.55, improving F1 from 0.851 to 0.864 ' |
| 'and precision from 0.905 to 0.934 while maintaining identical recall (0.803). This is a Pareto improvement: borderline false ' |
| 'positives are eliminated without losing any true positives. Above 0.85, recall degrades as the model becomes overly conservative.' |
| ) |
| pdf.fig(os.path.join(F, 'threshold_analysis.png'), |
| 'Fig. 9. Threshold analysis: (a) Precision/Recall/F1, (b) MCC. Optimal F1 at 0.55.', w=135) |
|
|
| pdf.sub('C.', 'Business Impact') |
| pdf.p( |
| 'Using average fraud amount $122.21 and $5/false alarm investigation cost, XGBoost yields the highest net savings ($6,936 on the ' |
| '42,559-transaction test set), catching 80.3% of fraud with only 6 false alarms ($30 cost). Logistic Regression catches 88.7% but ' |
| 'generates 1,229 false alarms ($6,145 cost), yielding only $1,554 net. The Autoencoder catches 100% but produces 21,209 false ' |
| 'alarms at $106,045 -- a net loss of $97,368. This underscores that maximizing recall alone is operationally counterproductive.' |
| ) |
| pdf.tbl( |
| ['Model', 'TP', 'FN', 'FP', 'Caught($)', 'FP Cost($)', 'Net($)'], |
| [ |
| ['XGBoost', '57', '14', '6', '6,966', '30', '6,936'], |
| ['Ensemble', '57', '14', '9', '6,966', '45', '6,921'], |
| ['LGBM-T', '58', '13', '24', '7,088', '120', '6,968'], |
| ['LR', '63', '8', '1229', '7,699', '6,145', '1,554'], |
| ['AE', '71', '0', '21209', '8,677', '106,045', '-97,368'], |
| ], |
| 'Table III: Business Impact Analysis' |
| ) |
|
|
| pdf.sub('D.', 'Explainability (SHAP and LIME)') |
| pdf.p( |
| 'SHAP analysis (2,000 test samples) reveals V4 (mean |SHAP|=1.913), V14 (1.843), and PCA_magnitude (1.113) as the dominant ' |
| 'fraud predictors. High V4 values push toward fraud; low (negative) V14 values are strongly associated with fraud, consistent ' |
| 'with EDA correlations. The engineered V10_V14_interaction ranks 9th, validating that interaction terms capture additional signal. ' |
| 'LIME analysis on a correctly classified fraud sample (P=1.0) shows Time_diff, V4, V12, and V14 as the strongest local contributors, ' |
| 'providing the granular instance-level explanation needed for regulatory compliance and analyst review.' |
| ) |
| pdf.fig(os.path.join(F, 'shap_summary.png'), |
| 'Fig. 10. SHAP summary: each dot = one sample; color = feature value; x-axis = SHAP impact on fraud prediction.', w=130) |
| pdf.fig(os.path.join(F, 'lime_explanation.png'), |
| 'Fig. 11. LIME explanation for a single fraud sample (P=1.0). Red = increases fraud risk; green = decreases it.', w=130) |
|
|
| |
| pdf.sec('VII', 'Error Analysis') |
| pdf.sub('A.', 'False Negatives (Missed Fraud)') |
| pdf.p( |
| 'XGBoost misses 14 of 71 fraud transactions (19.7%). Their mean predicted probability is only 0.013 -- the model is highly ' |
| 'confident they are legitimate, not borderline. Feature comparison explains why: FN transactions have V14 averaging -0.97 ' |
| 'vs -8.45 for true positives, V12 at -0.41 vs -7.69, and PCA_magnitude of 1.82 vs 12.25. These missed cases have feature ' |
| 'values dramatically closer to legitimate transactions, representing sophisticated fraud that mimics normal behavior. Lowering ' |
| 'the threshold would not help: at 0.12, only one additional FN would be caught while generating many more false alarms. ' |
| 'Catching these requires additional data sources (transaction sequences, device fingerprints, geography).' |
| ) |
| pdf.sub('B.', 'False Positives (False Alarms)') |
| pdf.p( |
| 'The 6 false positives have mean predicted probability 0.827 (some reaching 1.0). Their V14 averages -7.13 (vs -0.04 for TN) ' |
| 'and PCA_magnitude 7.86 (vs 0.28 for TN). These legitimate transactions genuinely exhibit fraud-like anomalous patterns -- ' |
| 'unusual but lawful spending (e.g., first-time purchases in unusual categories, international transactions). No model tuning can ' |
| 'distinguish these without additional contextual information.' |
| ) |
| pdf.sub('C.', 'Concept Drift and Retraining') |
| pdf.p( |
| 'Comparing model confidence between early and late test periods reveals a drift indicator of +0.115. We recommend: (1) weekly ' |
| 'PR-AUC monitoring on labeled data; (2) automated retraining when PR-AUC drops below 0.70; (3) sliding window training on 3-6 ' |
| 'months of recent data; (4) PSI monitoring on all features (alert when PSI > 0.25); (5) A/B testing for model updates; ' |
| '(6) quarterly fraud pattern reviews with domain experts.' |
| ) |
| pdf.fig(os.path.join(F, 'error_analysis.png'), |
| 'Fig. 12. Error analysis: (a) FN probability distribution, (b) FP probability distribution, (c) score distribution by class.', w=150) |
|
|
| |
| pdf.sec('VIII', 'Limitations') |
| pdf.bullet([ |
| 'PCA Anonymization: prevents domain-specific feature engineering (merchant, location, device) and limits interpretability.', |
| 'Temporal Scope: only two days of data, limiting drift assessment and seasonal pattern detection.', |
| 'Single-Institution: results may not generalize across banks, geographies, or payment networks.', |
| 'Static Features: no sequential transaction history (spending velocity, merchant novelty) which are critical in production.', |
| 'Static Threshold: optimal 0.55 was determined on test data and may shift; production needs dynamic adaptation.', |
| 'Simple Autoencoder: more advanced architectures (VAE, adversarial) might improve anomaly detection performance.', |
| ]) |
|
|
| |
| pdf.sec('IX', 'Future Work') |
| pdf.p( |
| 'Graph Neural Networks [19]: Modeling transaction networks as graphs enables fraud ring detection through suspicion propagation ' |
| 'across connected accounts -- impossible from individual transaction features alone.' |
| ) |
| pdf.p( |
| 'Real-Time Streaming: Integration with Apache Kafka and Flink would enable millions of transactions/second with consistent ' |
| 'sub-100ms latency guarantees. Federated Learning [20]: collaborative training across banks without sharing raw data preserves ' |
| 'privacy while expanding effective training sets for rare fraud types.' |
| ) |
| pdf.p( |
| 'LLM-Generated Explanations: Large language models could translate SHAP values into natural-language justifications for blocked ' |
| 'transactions, reducing analyst burden and satisfying regulatory requirements for explainable decisions.' |
| ) |
| pdf.p( |
| 'Temporal Sequence Modeling: Transformers or LSTMs on cardholder transaction sequences could capture behavioral patterns and flag ' |
| 'departures from established routines, treating fraud detection as time-series anomaly detection.' |
| ) |
|
|
| |
| pdf.sec('X', 'Conclusion') |
| pdf.p( |
| 'This paper presents a comprehensive fraud detection framework evaluating seven ML approaches on the European Cardholder benchmark. ' |
| 'XGBoost with cost-sensitive learning achieves best overall performance (PR-AUC 0.817, F1 0.851, MCC 0.852). Threshold optimization ' |
| 'to 0.55 improves F1 to 0.864 without sacrificing recall. Business impact analysis shows XGBoost catches 80.3% of fraud with only ' |
| '6 false alarms ($6,936 net savings), while the Autoencoder\'s 100% recall generates 21,000+ false alarms at $97,368 net loss.' |
| ) |
| pdf.p( |
| 'SHAP and LIME identify V4, V14, and PCA_magnitude as primary fraud discriminators. Error analysis reveals that 14 missed fraud ' |
| 'cases have feature profiles indistinguishable from legitimate transactions, requiring additional data sources to catch. The complete ' |
| 'system -- feature engineering, training, evaluation, explainability, and FastAPI deployment with sub-10ms latency -- demonstrates ' |
| 'that production-grade fraud detection is achievable with well-tuned classical ML. Tree-based ensembles, particularly XGBoost, ' |
| 'remain state-of-the-art for tabular fraud detection, outperforming deep learning and linear alternatives on all metrics that ' |
| 'matter for imbalanced classification.' |
| ) |
|
|
| |
| pdf.sec('', 'References') |
| refs = [ |
| '[1] A. Dal Pozzolo et al., "Calibrating probability with undersampling for unbalanced classification," IEEE CIDM, 2015.', |
| '[2] N. V. Chawla et al., "SMOTE: Synthetic Minority Over-sampling Technique," JAIR, vol. 16, pp. 321-357, 2002.', |
| '[3] A. Fernandez et al., Learning from Imbalanced Data Sets. Springer, 2018.', |
| '[4] T. Chen and C. Guestrin, "XGBoost: A scalable tree boosting system," ACM SIGKDD, 2016.', |
| '[5] G. Ke et al., "LightGBM: A highly efficient gradient boosting decision tree," NeurIPS, 2017.', |
| '[6] A. Pumsirirat and L. Yan, "Credit card fraud detection using deep learning," IJACSA, vol. 9, 2018.', |
| '[7] S. M. Lundberg and S.-I. Lee, "A unified approach to interpreting model predictions," NeurIPS, 2017.', |
| '[8] M. T. Ribeiro et al., "Why should I trust you?," ACM SIGKDD, 2016.', |
| '[9] R. Shwartz-Ziv and A. Armon, "Tabular data: Deep learning is not all you need," Info. Fusion, vol. 81, 2022.', |
| '[10] L. Grinsztajn et al., "Why do tree-based models still outperform deep learning on tabular data?," NeurIPS, 2022.', |
| '[11] T. Akiba et al., "Optuna: A next-generation hyperparameter optimization framework," ACM SIGKDD, 2019.', |
| '[12] R. J. Bolton and D. J. Hand, "Statistical fraud detection: A review," Statistical Science, vol. 17, 2002.', |
| '[13] Z. Zhang et al., "A model based on convolutional RNN for fraud detection," Complexity, 2021.', |
| '[14] A. A. Taha and S. J. Malebary, "An intelligent approach to credit card fraud detection," IEEE Access, vol. 8, 2020.', |
| '[15] V. Belle and I. Papantonis, "Principles and practice of explainable ML," Frontiers in Big Data, vol. 4, 2021.', |
| '[16] L. Prokhorenkova et al., "CatBoost: Unbiased boosting with categorical features," NeurIPS, 2018.', |
| '[17] S. Xuan et al., "Random forest for credit card fraud detection," IEEE ICNSC, 2018.', |
| '[18] T. Saito and M. Rehmsmeier, "The PR plot is more informative than ROC on imbalanced datasets," PLoS ONE, 2015.', |
| '[19] Y. Liu et al., "Pick and choose: A GNN-based imbalanced learning for fraud detection," Web Conf., 2021.', |
| '[20] Q. Yang et al., "Federated machine learning: Concept and applications," ACM TIST, vol. 10, 2019.', |
| '[21] Nilson Report, "Global card fraud losses," Issue 1209, 2022.', |
| '[22] A. Dal Pozzolo et al., "When is undersampling effective?," ECML PKDD, 2015.', |
| ] |
| pdf.set_font('Times', '', 7) |
| for ref in refs: |
| pdf.multi_cell(0, 3.2, ref) |
| pdf.ln(0.5) |
|
|
| out = os.path.join(PAPER_DIR, 'fraud_detection_paper.pdf') |
| pdf.output(out) |
| print(f"PDF saved: {out} ({pdf.page_no()} pages)") |
|
|
|
|
| if __name__ == '__main__': |
| build() |
|
|
