Fraud Detection System

Project Overview

A machine learning system engineered to detect fraudulent transactions in real-time, helping financial systems prevent fraud, minimize monetary loss, and enhance user trust.

Development Process

Preprocessed millions of transactions, handling massive class imbalance using Synthetic Minority Over-sampling Technique (SMOTE). Built and trained several models including Logistic Regression, Random Forests, and XGBoost, selecting the best model based on F1-score and Precision-Recall Area Under Curve. Wrapped the model in a Flask REST API for live inference scoring.

Key Features

• Advanced class-imbalance resolution utilizing SMOTE and undersampling techniques
• Feature engineering optimization including transaction-velocity tracking and geographic mapping
• Model benchmarking of RandomForest, XGBoost, and Logistic Regression models
• RESTful API integration via Flask for low-latency live transaction evaluation
• Detailed exploratory data analysis (EDA) highlighting fraud pattern clusters

Challenges & Solutions

Handling the severe class imbalance (less than 0.1% fraud cases) without overfitting the model. Solved by tuning probability thresholds, applying cross-validation on the SMOTE-augmented dataset, and prioritizing Precision-Recall curves over simple ROC-AUC.

Technologies Used