Add support for real-world datasets (Kaggle IEEE-CIS, Credit Card Fraud)

[stabrea]

Summary

The current model trains on synthetic or limited sample data. To validate real-world performance and establish honest benchmarks, we should integrate well-known public fraud detection datasets.

Motivation

• Real-world fraud data has very different class imbalance ratios (~0.17% fraud in IEEE-CIS)
• Benchmarking against known datasets allows comparison with published research
• Exposes edge cases that synthetic data misses (merchant category patterns, time-of-day effects)

Proposed Approach

1. Add data loader modules for:

   • Kaggle IEEE-CIS Fraud Detection (~590K transactions)
   • Kaggle Credit Card Fraud Detection (~284K transactions, PCA-transformed)

2. Preprocessing pipeline:

   • Handle missing values and categorical encoding for IEEE-CIS
   • Implement stratified train/test splitting preserving fraud ratio
   • Add feature engineering for temporal and aggregation features

3. Benchmarking:

   • Report precision, recall, F1, and AUC-PR (not just AUC-ROC, which is misleading with class imbalance)
   • Compare against published baselines
   • Document results in a benchmarks/ directory

Acceptance Criteria

• Data loaders for both datasets with automatic download/caching
• Preprocessing handles missing values and encoding
• Benchmark results documented with honest metrics
• README updated with dataset instructions and results table