Credit Fraud Detection — Class Imbalance at Scale

Role: Independent project · Period: Mar 2025 – Jun 2025

Overview

A focused study on what severe class imbalance does to fraud-detection model performance, and which combinations of resampling, dimensionality reduction, and model choice actually work in practice. The dataset: 200K+ credit transaction records, of which only ~500 were fraudulent — a ratio that breaks naive classifiers in well-known ways.

What I did

• Diagnosed the imbalance problem quantitatively, showing how default classifier outputs collapse onto the majority class without intervention and why standard accuracy is meaningless here.
• Applied multiple imbalance strategies in a controlled comparison: downsampling the majority class, class reweighting in the loss function, and PCA-based dimensionality reduction to surface signal that gets lost in high-dimensional fraud features.
• Benchmarked Random Forest, SGDClassifier, and MLP models on misclassification rate, false negatives (the costly error in fraud), and computational efficiency — because production fraud detection needs to be fast as well as accurate.
• Identified the optimal algorithm for large-scale fraud detection given the tradeoffs, with a clear writeup of why the winning combination beat the alternatives.

Why it matters

Class imbalance is the single most common reason ML models fail in fraud, medical, and security domains. This project is a clean case study in how to address it methodically — not just “throw SMOTE at it” but actually compare strategies on the metrics that matter for the business problem.

Tech stack

Python · scikit-learn · Random Forest · SGDClassifier · MLPClassifier · PCA · class reweighting · downsampling · cross-validation