Explainable AI for Fraud Detection: An Attention-Based Ensemble of CNNs, GNNs, and A Confidence-Driven Gating Mechanism

The model combines multiple AI approaches (CNNs, RNNs, GNNs, and LSTMs) to detect credit card fraud more accurately than single-method approaches, using an attention mechanism to weigh different predictive signals and select the most reliable predictions based on confidence levels.

Explainability is built into the fraud detection system through SHAP analysis, which identifies the top ten most important transaction features that distinguish fraud from legitimate activity, making it possible to understand why the system flags certain transactions rather than treating it as a "black box."

The confidence-based gating mechanism addresses a key responsible AI concern by automatically selecting which prediction method to trust, reducing uncertainty in high-stakes financial decisions where false positives (blocking legitimate transactions) and false negatives (missing fraud) both have real consequences for users and institutions.

Testing on multiple datasets demonstrates the method's robustness, suggesting the approach generalizes beyond a single fraud detection scenario, which is important for responsible deployment across different financial institutions and transaction types.

The focus on both accuracy and explainability reflects growing recognition that fraud detection systems must be trustworthy and interpretable, particularly since these systems directly impact customer trust in digital financial services and can affect vulnerable populations disproportionately if they produce biased results.