Fraud Detection Pattern

A machine-learning–driven system designed to identify abnormal behavioral patterns and flag potentially fraudulent activity.
The project focuses on data preprocessing, feature engineering, model training, evaluation, and real-time prediction capability.

---

Highlights

• 95.34% accuracy on a labelled dataset of 10,000 transactions (test split).
• Graph visualization of transaction networks (Neo4j + Neovis.js) to explore suspicious clusters and entity relationships.

---

Features

• Transaction graph visualization using Neo4j for relationship-aware investigation
• Preprocessing pipeline for cleaning, encoding, and transforming raw data
• Feature engineering to extract fraud indicators (behavioral & network features)
• Model training with multiple algorithms (e.g., Random Forest, XGBoost, Isolation Forest)
• Evaluation using precision, recall, F1-score, ROC-AUC, plus overall accuracy
• Exportable model for batch and realtime inference
• Modular structure for easy experimentation and extension

---

How It Works

1. Load and clean the input dataset
2. Engineer features that capture transactional behavior
3. Train fraud-detection models
4. Evaluate model performance with fraud-sensitive metrics
5. Deploy the model for prediction on new data

---