Project Record
Card Transaction Fraud Detection
Explored anomaly and supervised models on a severely imbalanced card-transaction dataset to rank transactions for analyst review.
Executive Summary
Card fraud detection is a severe class-imbalance problem: fraud is rare, labels arrive late, and analysts can only review a fixed number of transactions per day. The useful question is not overall accuracy but whether ranked scores surface real fraud cases early. I worked through an obfuscated retail transaction dataset to compare anomaly detection and supervised models under these operational constraints.
A balanced ensemble gave the best trade-off for review workflows: high precision on flagged cases and useful ranking when only the top few hundred transactions can be checked daily. The takeaway was operational—how to score and queue under imbalance—not a claim of production-ready deployment.
This demonstrates imbalanced classification, feature engineering, precision–recall evaluation, and fraud analytics workflow design.
Problem
Given transaction metadata (account, merchant, amount, POS mode, geography, timing), flag likely fraud without drowning reviewers in false positives. At sub-1% fraud prevalence, accuracy is a poor target; the useful question is whether ranked scores surface real cases early.
Approach
- Merged delayed fraud labels onto the transaction feed and explored class balance, amount patterns, and feature correlations.
- Engineered simple behavioural signals from timestamps and spend relative to available balance.
- Compared several classifiers using precision–recall AUC, which is better suited to rare positives than raw accuracy.
- Evaluated both unsupervised anomaly detection (Isolation Forest) and supervised ensembles with class reweighting.
Outcome
A balanced ensemble gave the best trade-off for review workflows: high precision on flagged cases, moderate recall on known fraud, and useful ranking when only the top few hundred transactions can be checked each day. The main takeaway was operational—how to score and queue under imbalance—rather than a production-ready detector.
Trade-offs
Late labels blur the training boundary; high-cardinality merchant and account IDs do not generalize cleanly; and precision-first scoring naturally leaves recall on the table unless paired with rules or graph signals elsewhere in the stack.