Afolabi Olawale

Menu

Close

Fighting Fraud with Machine Learning
Built a production-grade fraud detection system on 1M+ interbank transactions with a 0.3% fraud rate — no synthetic balancing tricks. Engineered 38 behavioral features, trained interpretable models (Logistic Regression → Random Forest), and used SHAP to make every flagged transaction explainable to analysts and regulators.
Visit WebsiteGithub
Fraud Detection
Random Forest
SHAP
Feature Engineering
Python
Scikit-learn
Imbalanced Learning
Image of Fighting Fraud with Machine Learning

#Fighting Fraud with Machine Learning

This isn't a Kaggle notebook with synthetic oversampling. It's the messy, iterative work of building fraud detection that actually holds up when fraudsters adapt.

#The Problem

1 million transactions. 3,000 fraudulent. That's a 0.3% fraud rate across multiple Nigerian banks — the kind of brutal class imbalance you get in production. Miss fraud and it costs real money. Flag too many false positives and you destroy user trust.

#What I Actually Built

#Feature Engineering (Where 80% of the Work Happened)

Started with 62 features. Aggressively pruned to 17 that actually matter. Every feature answers one question: "How weird is this transaction for this user?"

  • Temporal encodings — fraudsters are creatures of habit (midnight to 5 AM spikes confirmed)
  • Velocity signals — transaction count and amount over 24h/7d/lifetime windows
  • Behavioral diversity — channel switching and location hopping patterns
  • Risk composites — weighted combinations encoding multi-dimensional suspicion

#Models & Honest Results

Logistic Regression (Baseline): AUC-ROC 0.70 — caught 54% of fraud but with a 99% false positive rate. Completely unusable.

Random Forest (Production Model): AUC-ROC 0.82 — catches nearly 2 out of 3 fraud cases. Got there through careful depth limiting, balanced class weights, and letting the trees find the non-linear boundaries that logistic regression missed.

#SHAP Interpretability

"The model says so" doesn't fly with fraud analysts or regulators. Built full SHAP analysis — summary plots, dependence plots, force plots, waterfall breakdowns. Transaction amount drives 52% of importance, followed by the composite risk score I engineered at 18%.

#Key Findings That Held Up
  • Time-of-day is real — fraud risk spikes 2-3x between midnight and 5 AM
  • Mobile is the wild west — faster fraud velocity, higher amounts, weaker controls
  • The best fraud signal: "doing something you've never done before, quickly"
  • Location alone is weak. Location + velocity + unusual channel = strong signal

#Why This Matters

Fraud detection is adversarial. Fraudsters test your boundaries, find blind spots, and adapt. This project treats it that way — no magic accuracy claims, honest about trade-offs, built for the people who actually need to trust and use the model.