Machine Learning for Fraud Detection in E-commerce Transactions

Overview

This project investigates the application of machine learning techniques to enhance fraud detection in e-commerce transactions. By leveraging a comprehensive dataset from Vesta, we explore feature engineering, distance prediction, and clustering analysis to identify fraudulent activities.

Problem Statement

The increasing sophistication of financial fraud poses significant challenges to businesses and consumers. Traditional rule-based fraud detection systems often struggle to keep pace with evolving fraudulent tactics. This project aims to develop more robust and accurate fraud detection models using machine learning.

Methodology

This project addresses three key research questions:

RQ1: Feature Engineering and Selection

• Objective: Improve fraud detection accuracy by identifying and engineering the most predictive features.
• Techniques: Recursive Feature Elimination (RFE), Feature Importance from Gradient Boosting, Principal Component Analysis (PCA).

RQ2: Predicting Transaction Distances

• Objective: Develop models to predict transaction distances and identify geographic anomalies indicative of fraud.
• Techniques: Linear Regression, XGBoost.

RQ3: Clustering for Coordinated Fraud Detection

• Objective: Utilize clustering techniques to uncover groups of transactions potentially associated with coordinated fraud.
• Techniques: K-Means Clustering, HDBSCAN, Hierarchical Clustering.

Results

• Feature Engineering: PCA significantly enhanced model accuracy, highlighting its effectiveness in capturing relevant data structures.
• Distance Prediction: XGBoost models demonstrated promising results in predicting transaction distances, aiding in the identification of high-risk transactions.
• Clustering Analysis: K-Means Clustering provided the most interpretable and well-separated clusters, potentially revealing patterns of coordinated fraud.

Data

• Source: "IEEE-CIS Fraud Detection" dataset from Kaggle, provided by Vesta.
• Size: Over 140,000 transactions with 434 features (transaction details, card information, addresses, Vesta-engineered features).

Contact

Cem Kazan - kzncem@gmail.com# fraud-detection-ml