Tabular ML
IEEE-CIS Fraud Detection
5
model families behind one factory: LR, RF, LightGBM, XGBoost, CatBoost
Real transaction data with heavy missingness and severe class imbalance, organised as a reproducible pipeline instead of one notebook.
What I did
- Left-joined identity data onto transactions by TransactionID, added frequency encoding for high-cardinality card fields and time features from TransactionDT.
- Evaluated by ROC-AUC because of the 3 to 4% fraud rate, with LightGBM as the main model.
Stack
LightGBMXGBoostCatBoostscikit-learn