Project Overview

An end-to-end fraud detection pipeline built on the IEEE-CIS Fraud Detection Kaggle competition dataset. This repository demonstrates a complete machine learning workflow for detecting fraudulent transactions in real-time financial systems. Starting with 590K+ transactions and 434 raw features, we engineer intelligent risk signals, benchmark multiple gradient boosting algorithms, and maintain a rigorous experiment log — all designed for production-ready deployment.

What Makes This Different?

Feature Engineering Guess & pray Hypothesis-driven with adversarial validation Model Selection Single algorithm LGBM · XGBoost · CatBoost benchmark Validation Naïve train/test Stratified splits + early stopping + threshold tuning Experiment Tracking Scattered notebooks Versioned experiment log with F1 deltas Categorical Handling One-hot explosion (485+ features) Native LightGBM category support Model Analysis Basic accuracy SHAP · Calibration · PR Curves · Confusion Matrix

Live Experiment Dashboard

0.84370.25CatBoost442+40.1% over baseline

The project utilizes advanced machine learning techniques to predict customer risk levels. Key features include:

• Comprehensive Feature Engineering: Developed over 400 features, including time-based, count-based, and identity-based features, to capture complex fraud patterns.

• Model Benchmarking: Evaluated and compared multiple gradient boosting models (LightGBM, XGBoost, CatBoost) to identify the best performing algorithm.

• Rigorous Validation: Employed stratified splits, early stopping, and threshold tuning to ensure robust model performance.

• Explainable AI (XAI): Integrated SHAP values for model interpretability, allowing for detailed analysis of individual predictions and feature importance.

• Production-Ready Pipeline: Designed for real-time deployment, with optimized data processing and model serving capabilities.

Engineered Features

The project focuses on creating meaningful features that capture user behavior and transaction anomalies. Notable engineered features include:

• UID-based features: Aggregations and statistics derived from user identifiers (e.g., uid_transaction_count, uid_mean_amount) to identify suspicious patterns.

• Temporal features: Analysis of transaction timing (e.g., uid_mean_hour) to detect unusual activity.

• Match indicators: Features capturing consistency between different data points (e.g., card address vs. billing address).

Model Performance & Visualizations

The project provides extensive visualizations and performance metrics to demonstrate the effectiveness of the fraud detection models. This includes:

• Confusion Matrix Analysis: Visualizes true positives, true negatives, false positives, and false negatives to assess classification accuracy.

• Precision-Recall Curve: Illustrates the trade-off between precision and recall, crucial for imbalanced datasets.

• Model Calibration: Assesses how well the predicted probabilities align with actual fraud rates.

• Feature Importance: Identifies the most influential features driving fraud predictions across different models.

• SHAP Analysis: Provides both global and local explanations for model predictions, enhancing transparency and trust.