Fraud Detection System Capstone Project
Designed and built a production-grade fraud detection pipeline on 5M+ transactions, identifying behavioural, temporal, and device-based risk patterns.
CONTEXT
Financial transaction data · Highly imbalanced (~3–4% fraud rate)
TOOLS
Python, XGBoost, LightGBM, MLflow, SHAP
AUC-PR optimised
Cost-based threshold tuning
IMPACT
-
check_circle
Identified transaction-level risk signals (velocity, behavioural deviation, device novelty)
-
check_circle
Reduced false positives through threshold optimisation aligned with business cost
-
check_circle
Improved interpretability using SHAP for auditability
01 · FEATURE ENGINEERING
Engineered behavioural, temporal, and device-level features to capture fraud patterns and anomalies.
02 · MODELLING
Trained XGBoost and LightGBM models with focus on AUC-PR for imbalanced classification.
03 · EVALUATION
Ensemble of XGBoost & Random Forest. Bayesian optimisation. Precision-recall analysis, confusion matrix diagnostics, and cost-based performance evaluation.
