Work samples with evidence attached.

E-commerce Purchase Behavior Prediction

Problem: The data contains 599,116 user-item pairs with only 1.12% purchase positives, so accuracy alone would hide missed purchases.

Method: I built leakage-safe behavior sequence features, category target encoding, user-category z-scores, time-window density features, and an XGBoost model with imbalance-aware training.

Result: The baseline reached AUC-ROC 0.9994, AUC-PR 0.9760, and F1 0.9639, with SHAP and error analysis used to explain model behavior.