Machine Learning - Santander Transaction Prediction (Final Project)

Final Project (Grade 95, M.Sc. Data Science, HIT). Binary classification of customer transactions on the Santander Kaggle competition — 200,000 records, 200 anonymized features, 9:1 class imbalance (hence AUC-ROC as the target metric).

Key Features

• Feature Engineering at Scale: 217 statistical features engineered on top of the 200 anonymized originals.
• Feature Selection: Importance-based narrowing to the 65 most impactful — a leaner model that filters statistical noise.
• Model Comparison: CatBoost vs. Logistic Regression baseline, with train/validation/test gap tracking against overfitting.
• External Validation: Final model submitted to Kaggle and scored on the competition's unseen test set.

Results

AUC 0.894 validation / 0.859 Kaggle private leaderboard — a modest 0.035 generalization gap on the competition's fully unseen test set.

Repository Structure

• Santander_Transaction_Prediction_Final.ipynb: Full implementation — EDA, feature engineering, selection, modeling, Kaggle submission (Hebrew narrative; code and charts are language-independent).
• Final_Project_Instructions.pdf: Original project instructions.
• train.csv / test.csv: The official competition datasets (via Git LFS).
• sample_submission.csv / submission_santander.csv: Submission template and final predictions.
• kaggle_score.png: Official Kaggle leaderboard score screenshot.