Predictive Analytics • 100K transactions

Deep Learning-Based Credit Card Fraud Detection

Fraud analytics solution with threshold tuning to balance detection performance and customer experience.

PythonScikit-learnTensorFlow/KerasPyTorch GeometricKeras Tuner

Deep Learning-Based Credit Card Fraud Detection cover

Executive summary

Designed a fraud detection analytics solution with threshold tuning to balance detection performance and customer experience, translating business requirements into measurable evaluation decisions.

Problem

Fraud detection must balance false positives (customer friction) vs false negatives (loss).
Success depends on the right metric + threshold, not only model accuracy.

Data

Scale

100K transactions

Notes

Fraud problems are typically imbalanced; PR-focused evaluation is often more informative (add your chosen metrics).

Approach

Built and compared multiple ML/DL approaches (update exact models later).
Evaluated trade-offs using appropriate metrics for imbalance (update with your final choice).
Optimized the decision threshold to align with operational/business trade-offs.

Insights

Threshold selection materially changes operational outcomes (fraud captured vs customer friction).
PR-focused analysis provides clearer signal on rare-event detection (add your plots).

Impact

Delivered decision-threshold guidance for deployment-focused operations.
Created an evaluation framework to support cost-aware risk decisions.

Screenshots

Fraud detection - PR curve or evaluation plot — Dataset / class imbalance summary (the key chart/table showing fraud vs non-fraud distribution).

Fraud detection - model comparison table — Transformer Model performance curve (PR curve and ROC curve).