When fraud is rare,
accuracy isn't enough.
In large-scale e-commerce transaction streams, fraudulent payments represent less than 3.5% of volume. Standard classifiers achieve near-perfect accuracy by blindly predicting every charge as legitimate—while failing 100% of financial crime investigations.
Explore Research by Focus Area
Each module provides focused interactive scientific tools, verified empirical data, and real-time computation.
Interactive Research Pipeline
Follow the 8 rigorous stages of the thesis methodology. Toggle between training and real-time inference workflows to inspect anti-leakage safeguards, latent manifold oversampling, and model scoring.
Dataset Ingestion & Merging
Merge 590,540 e-commerce transaction records with identity verification tables on TransactionID.
Model Comparison Workbench
Comprehensive evaluation of all 6 experimental configurations on the 118,108 held-out test transactions. Average Precision (PR-AUC) serves as the primary metric, isolating the contribution of Autoencoder latent-space synthesis and Optuna Bayesian optimization.
| Scenario / Method | Oversampling Technique | PR-AUC (AP) | Δ vs Base | ROC-AUC | Precision | Recall | F1-Score | Threshold |
|---|---|---|---|---|---|---|---|---|
| Baseline LightGBM | None | 0.810074 | — | 0.966358 | 83.37% | 70.23% | 0.7624 | 0.73 |
| Baseline + Optuna | None | 0.854378 | +0.044304 (+5.5%) | 0.969428 | 91.29% | 72.58% | 0.8087 | 0.56 |
| SMOTE + LightGBM | SMOTE (Feature Space) | 0.824418 | +0.014344 (+1.8%) | 0.966620 | 86.16% | 70.96% | 0.7782 | 0.23 |
| SMOTE + Optuna | SMOTE (Feature Space) | 0.854120 | +0.044046 (+5.4%) | 0.969057 | 89.17% | 74.35% | 0.8109 | 0.14 |
| AE Latent OversamplingProposed | Autoencoder Latent Interpolation | 0.827153 | +0.017080 (+2.1%) | 0.967231 | 87.17% | 70.52% | 0.7797 | 0.23 |
| AE Latent + Optuna (Proposed)Proposed | Autoencoder Latent Interpolation | 0.856187 | +0.046114 (+5.7%) | 0.970334 | 89.91% | 73.77% | 0.8104 | 0.16 |
Precision-Recall & ROC Evaluation Curves
Hover or drag across curves to inspect dynamic decision threshold trade-offs on the 118,108 held-out test transactions.
Autoencoder Sensitivity Grid (Validation Split)
Grid search across 4 latent dimensions and 5 target fraud rates using default LightGBM. Locked at latent_dim=8 and rate=10% (PR-AUC: 0.784424).
| Latent Dim (Z) \ Target Rate | 10% Fraud | 15% Fraud | 20% Fraud | 30% Fraud | 50% Fraud |
|---|---|---|---|---|---|
| dim = 8 | 0.784424 | 0.784312 | 0.782129 | 0.782037 | 0.777077 |
| dim = 16 | 0.782806 | 0.780727 | 0.781784 | 0.778684 | 0.776781 |
| dim = 32 | 0.783631 | 0.783258 | 0.779935 | 0.778753 | 0.777079 |
| dim = 64 | 0.783688 | 0.782399 | 0.780670 | 0.778845 | 0.778310 |
Research Finding: Compacting fraud features into 8 dimensions avoids noise in higher dimensions, while an oversampling target of 10% provides sufficient representation density without overwhelming normal payment transaction geometry.
Interactive Decision Threshold Explorer
In fraud prevention, raw model probability is not a binary decision. Explore how sliding the classification threshold navigates the critical operational trade-off between financial chargeback losses and customer checkout friction.
Dynamic Confusion Matrix
N = 118,108Ground truth vs model predictions at active threshold τ.
Transaction Investigation & WASM Simulator
Examine authentic evaluation cases with precomputed TreeSHAP force plots, or test arbitrary custom transaction attributes using the in-browser LightGBM WebAssembly inference engine.
Transaction #3184276
Actual Ground Truth: Confirmed FraudHigh-Velocity High-Value Fraud (True Positive)
Methodology Notes & Constraints
Maintaining scientific fidelity is paramount. Below are the architectural justifications, mathematical rationale, and explicit limitations of the thesis research.
Autoencoder is NOT an Inference-Time Detector
A common misconception in applied fraud detection is treating the Autoencoder reconstruction error as a live anomaly score. In this methodology, the undercomplete Autoencoder (432 → 128 → 64 → 8) is used strictly during offline training to construct a smooth, low-dimensional latent manifold for minority-class SMOTE interpolation. Real-time inference probability scores are computed exclusively by the fitted LightGBM classifier.
Strict Train-Only Preprocessing Fit
To prevent lookahead data leakage:
- Categorical frequency tables for 49 IEEE-CIS features were computed solely from the 354,324 training rows.
- Numerical median imputations for 383 features were fitted strictly on train data.
- Z-score standardization means and standard deviations were derived exclusively from the training split and clipped to [-10, 10].
Why Average Precision (PR-AUC) is Primary
With 3.50% fraud prevalence, ROC-AUC gives an overly optimistic assessment because the large number of true negatives (113,976) depresses the False Positive Rate (FPR = FP / [FP + TN]). As proven by Davis & Goadrich (2006) and Saito & Rehmsmeier (2015), Precision-Recall AUC (Average Precision) directly isolates positive class trade-offs without dilution by the vast legitimate majority.
Real-World Deployment Constraints
1. Concept Drift: The IEEE-CIS benchmark reflects historical transactions over a specific time window. Production fraud detection requires continuous model retraining and online drift detection.
2. Static Artifact Exclusions: Large serialized model binaries (.booster files) are deliberately excluded from public git versioning to maintain repository hygiene.
Reproduce on Kaggle with One Click
Run the locked script kaggle/ieee_final_oversampling_kaggle.py against the official IEEE-CIS competition input.