GradCFA: A Hybrid Gradient-Based Counterfactual and Feature Attribution Explanation Algorithm for Local Interpretation of Neural Networks

Hybrid explanations provide more complete transparency: GradCFA combines two complementary explanation approaches—counterfactual explanations (showing what would need to change for a different outcome) and feature attribution (showing which inputs most influenced the decision)—offering users a more comprehensive understanding of AI decisions than either method alone.

New method works for complex, multi-class decisions: While previous counterfactual explanation research focused mainly on simple yes/no decisions, GradCFA extends to multi-class scenarios, making it applicable to real-world AI systems in healthcare and finance that must choose among many possible outcomes rather than just two.

Balances three critical quality standards simultaneously: The framework explicitly optimizes for feasibility (explanations are realistic), plausibility (explanations make sense in context), and diversity (multiple different explanations are provided), addressing a key limitation where existing methods often excel at one quality while sacrificing others.

Validated against leading competitors: Testing against established methods like SHAP, Wachter, and DiCE demonstrates that GradCFA generates explanations that are both feasible and diverse while maintaining the ability to identify which features actually influenced the AI's decision.

Advances accountability in high-stakes domains: By improving how AI systems explain their decisions in critical fields like healthcare and finance, this work supports more responsible AI deployment by enabling stakeholders to verify that systems are making decisions for the right reasons.