Explainable Artificial Intelligence and Causal Inference based ATM Fraud Detection

Explainability matters for financial trust: The paper emphasizes that banks must prioritize interpretable AI models for fraud detection because customers need to understand why transactions are flagged or blocked. Decision Trees achieved nearly identical performance (0.958 AUC) to the best-performing model while offering superior interpretability, demonstrating that explainability doesn't require sacrificing accuracy.

Imbalanced fraud data requires careful handling: ATM fraud datasets are inherently skewed since fraudulent transactions are rare. The study found that SMOTE (Synthetic Minority Oversampling Technique) and its variants significantly improved model performance, suggesting that responsible fraud detection requires deliberate technical approaches to prevent bias against legitimate transactions.

Simpler models can outperform complex ones: While advanced techniques like Gradient Boosting Trees achieved the highest AUC (0.963), the authors highlight that Decision Trees provide comparable results with lower computational complexity. This finding supports responsible AI practices by showing that organizations don't always need "black box" systems to achieve effective fraud detection.

One-class classification offers an alternative approach: Beyond traditional binary classification, the paper explores One-Class Classification methods (like IForest), which may better reflect real-world fraud scenarios where fraudulent patterns are fundamentally different from normal transactions, potentially reducing false positives that frustrate legitimate customers.

Context-specific solutions are necessary: The study uses an ATM transaction dataset from India, highlighting that fraud detection systems must be tailored to regional patterns and banking practices rather than relying on generic global models.