Deep Learning for Real-Time Credit Card Fraud Detection

Abstract

The presentation will cover three main areas:

• High-Performance Deep Learning: I'll detail the use of a feedforward deep neural network (DNN) that significantly outperforms traditional machine learning models (like Random Forest and Logistic Regression) in detecting subtle, non-linear fraud patterns. Our model achieved 96.2% accuracy and a crucial 97.5% recall.
• Opening the Black Box with SHAP: I will explain how SHapley Additive exPlanations (SHAP) is used as a post-hoc technique to provide human-interpretable explanations for every single prediction. This addresses the "black box" problem, which is vital for regulatory compliance (like the "right to explanation") and customer trust in finance.
• Actionable Insights for Analysts: Attendees will learn how this combined system offers both global and local explainability. I'll demonstrate how: feature importance plots (global view) validate that Transaction Amount and Time of Day are the strongest risk drivers; and instance-level SHAP analysis (local view) reveals the dynamic, context-specific reasons for a fraud score (e.g., a "New Device ID" being the critical factor). Attendees will benefit by understanding how to implement a system that achieves both state-of-the-art predictive performance and the necessary transparency, trust, and accountability for high-risk applications in digital finance