🕵️ Unsupervised Credit Card Fraud Detection

📌 Problem Statement

Credit card fraud is a major issue for banks and payment systems. Fraudulent transactions are rare (~0.17% in this dataset), making detection difficult. Traditional supervised learning requires labeled fraud data, which is often unavailable in real-world scenarios.
This project applies unsupervised anomaly detection to identify suspicious transactions without relying on labels.

---

🎯 Objectives

• Detect fraudulent transactions using unsupervised ML techniques.
• Avoid data leakage by using time-based splits.
• Build a beginner-friendly yet resume-worthy project that can later scale into an industry-ready solution.

---

📊 Dataset

• Source: Kaggle European Credit Card Transactions Dataset
• ~284,807 transactions
• Features: 28 anonymized PCA features (V1–V28), Time, Amount, Class (0=Normal, 1=Fraud).
• Extremely imbalanced (frauds ≈ 0.17%).

---

⚙️ Tech Stack

• Python 3
• Google Colab (for development)
• Libraries:
  • pandas, numpy → data manipulation
  • matplotlib, seaborn → visualization
  • scikit-learn → Isolation Forest, Local Outlier Factor
  • umap-learn → dimensionality reduction
  • gradio → tiny interactive demo

---

🔍 Approach

1. Data Loading: Kaggle dataset imported via Colab/Drive.
2. Minimal EDA: Checked class imbalance, transaction time, and fraud percentages.
3. Time-based Split: Train = first 70%, Test = last 30% → prevents leakage.
4. Downsampling: Kept all fraud rows, sampled normal rows for faster training.
5. Feature Engineering:
   • Used anonymized V1–V28
   • Robust-scaled Amount
   • Added simple time features (hour, night_flag)
6. Models:
   • Rule-based baseline (night + high amount)
   • Isolation Forest (main unsupervised model)
   • Local Outlier Factor (LOF) for comparison
7. Evaluation:
   • Precision@K (top-K suspicious transactions)
   • Recall@K
   • PR-AUC (average precision)
8. Gradio Demo: Simple app where user inputs amount/time → model outputs fraud likelihood.

---

📈 Results (Mini Project)

• Isolation Forest detected fraud transactions far above random chance.
• PR-AUC ≈ 0.87 (much better than random baseline).
• Showed top 20 suspicious transactions for analysts.

(Insert a plot or table screenshot here → place in reports/ and link)

---

🚀 Future Improvements (Scaling to Major Project)

• Hybrid models (Autoencoder + Isolation Forest).
• Explainability with SHAP/LIME.
• Real-time fraud detection API (FastAPI + Docker).
• Streaming pipeline (Kafka + Spark/Flink).
• Graph-based fraud networks (Neo4j, PyTorch Geometric).
• Deployment to cloud (AWS/GCP/Azure).

---

📂 Repository Structure

Fraud-detection/ │── README.md │── requirements.txt │── .gitignore │── Fraud_Detection_Mini_Project.ipynb │── reports/ │ ├── Fraud_vs_Normal.png │ └── Demo_Gradio.png