Load Kaggle CC Fraud Dataset (284,807 European transactions)

Exploratory Data Analysis — verify 0% missing values, plot distributions

Preprocess Data — log-transform Amount, z-score normalise Amount & Time

Feature Selection — choose FSS0 / FSS1 / FSS2 strategy

80/20 Stratified Train–Test Split — preserving 0.172% fraud rate in test set

Apply SMOTE (training set only) — balance to 1:1 ratio, 454,902 records

Train 24 model configurations — 4 MT × 3 FSS × 2 RT

Evaluate all configurations — F1, ROC-AUC, Precision, Recall, Confusion Matrix

Apply SHAP + LIME to best model — generate EU AI Act compliance documentation