Explainable Fraud Detection with Deep Symbolic Classification

Explainability is critical for fraud detection: The paper emphasizes that fraud detection models must provide clear explanations for their decisions because customers may dispute them and regulators require transparency to build trust—a requirement that traditional "black box" AI models struggle to meet.

Deep Symbolic Classification produces human-readable rules: Unlike complex neural networks, DSC generates fraud detection decisions as concise mathematical expressions in closed form, making both individual decisions and the overall model logic understandable to non-technical stakeholders and regulators.

Class imbalance is a major practical challenge: Fraud detection datasets contain "several orders of magnitude more legitimate transactions than fraudulent ones," which can cause standard AI models to perform poorly; DSC addresses this by directly optimizing for metrics like F1 score that remain reliable despite severe class imbalance.

Data-driven approaches can improve fraud detection: The paper demonstrates that combining large historical datasets with machine learning can better capture the "noisy, dynamic nature of fraud" compared to rule-based systems alone, while maintaining the explainability that business and regulatory stakeholders require.

Responsible AI requires balancing multiple objectives: This work shows that effective fraud detection systems must simultaneously achieve accuracy, explainability, fairness (handling class imbalance), and regulatory compliance—not just optimize for a single performance metric.