Transaction Fraud Detection ML Pipeline

Role: Volunteer Data Scientist — Google Engineering Team · Period: Winter 2023

Overview

Engineered a production-grade fraud detection pipeline over 800K+ Google Pay transactions, working alongside Google engineers in a volunteer capacity. The deliverable was a tuned Gradient Boosting model with stable performance across train/test splits — ready for deployment to a real fraud-prevention workflow.

What I did

• Built the full pipeline: outlier detection, categorical encoding, feature engineering, and dataset splitting strategies designed for the heavy class imbalance typical of fraud detection (frauds are <1% of transactions).
• Trained and benchmarked Logistic Regression, GBDT, and MLP models with cross-validation and systematic hyperparameter tuning, comparing how each handled the precision–recall tradeoff that defines fraud-prevention systems.
• Reduced false positives by 8.7% and improved recall by 2.87% in the final model versus baseline — meaningful in production because false positives directly translate to legitimate customers being blocked.
• Delivered a production-ready GBDT model with stable train/test performance, suitable for deployment without additional retraining.

Why it matters

Fraud detection lives or dies on the precision–recall tradeoff. Pushing recall up without driving false positives through the roof is the whole game, and the 8.7% reduction in false positives at slightly higher recall is exactly the kind of asymmetric improvement that fraud teams care about — it means fewer angry customers and more caught fraud, simultaneously.

Tech stack

Python · scikit-learn · Logistic Regression · Gradient Boosting (GBDT) · MLP · cross-validation · hyperparameter tuning · feature engineering