🛡️ Fraud Detection — Stacking-Based Machine Learning System

An end-to-end fraud detection project built on the IEEE-CIS Fraud Detection dataset, using multiple powerful base models, advanced preprocessing, and a meta-learner (stacking ensemble) to achieve strong performance on highly imbalanced transaction data.

This repository focuses on:

• Realistic time-based splits
• Out-of-fold (OOF) predictions
• Optuna hyperparameter optimization
• SHAP-based explainability
• Ensemble learning for robust fraud detection

---

✨ Project Highlights

• 🔹 Designed for extreme class imbalance (fraud < 1%)
• 🔹 Multiple strong base models (LightGBM, CatBoost, Isolation Forest, Autoencoder)
• 🔹 Stacking / Meta-learning using XGBoost
• 🔹 Careful data leakage prevention via real-time splits
• 🔹 Model explainability using SHAP
• 🔹 Notebook-based, experiment-driven workflow

---

🧠 High-Level Approach

Fraud detection is not just classification — it is risk ranking under asymmetric cost.

This project follows a stacking ensemble strategy:

1. Preprocess raw transaction data
2. Train multiple base models using out-of-fold predictions
3. Use base model outputs as features
4. Train a meta-learner to combine model strengths
5. Explain predictions using SHAP values

This approach improves generalization and captures different fraud patterns that a single model may miss.

---

📊 Dataset 📦 IEEE-CIS Fraud Detection Dataset

▶️ Recommended Execution Order

1️⃣ 01_preprocessing_real_time_split.ipynb

2️⃣ 01b_preprocessing_catboost_dataset.ipynb

3️⃣ Base Models

02_lightgbm_base_model_oof_optuna_shap.ipynb

03_isolation_forest_base_model_oof_optuna_shap.ipynb

04_autoencoder_base_model_oof_optuna_shap_v2.ipynb

05_catboost_base_model_oof_optuna_shap_fast.ipynb

4️⃣ Meta Learner

06_xgb_meta_learner_optuna_shap_with_cat.ipynb

Licensed under the MIT License See LICENSE for details.

👤 Author

Akshat Fraud Detection & Machine Learning GitHub: https://github.com/Dhingraakshat

🗂️ Repository Structure

Fraud-detection/
│
├── LICENSE
├── README.md
│
├── Dataset/
│   └── IEEE_CIS/
│       └── (IEEE-CIS Fraud Detection dataset files)
│
├── Preprocessing/
│   ├── 01_preprocessing_real_time_split.ipynb
│   └── 01b_preprocessing_catboost_dataset.ipynb
│
├── Base_Models/
│   ├── 02_lightgbm_base_model_oof_optuna_shap.ipynb
│   ├── 03_isolation_forest_base_model_oof_optuna_shap.ipynb
│   ├── 04_autoencoder_base_model_oof_optuna_shap_v2.ipynb
│   └── 05_catboost_base_model_oof_optuna_shap_fast.ipynb
│
└── Meta_Learner/
    └── 06_xgb_meta_learner_optuna_shap_with_cat.ipynb