This project explores how different machine learning models can help detect credit card fraud, a costly problem for banks and customers.
I experimented with both unsupervised and supervised techniques to spot fraudulent transactions based on transformed features and a binary target variable (0 = non-fraudulent, 1 = fraudulent).
- Isolation Forest: detects outliers by making random splits and isolating anomalies
- Autoencoder: learns to reconstruct normal transactions and flags anomalies with reconstruction loss.
- Random Forest: tested on original features, features combined with anomaly scores, and autoencoder embeddings.
- Multilayer Perceptron (MLP): the trained neural network has a simple architecture.
- Best unsupervised model: Autoencoder (AUC = 0.73)
- Best supervised model: MLP (AUC = 0.95)
Autoencoder had quite quite impressive performance, given that it does not require labels. This is efficient in real-life cases as well, when labeled data is absent. When labels are available, supervised learning models are far more effective at catching fraud.
An interesting next step would be to explore more systematic hyperparameter tuning, which may further improve the models' performance .