Explainable AI Through a Democratic Lens: DhondtXAI for D'Hondt-Projected Feature Attribution

Democratic principles can improve AI explanations: DhondtXAI applies the D'Hondt rule—a proportional representation method used in parliaments worldwide—to explain AI decisions. This approach treats features like political parties competing for "seats" in the explanation, ensuring fair allocation of credit based on their actual contribution to the model's output.

Separating positive and negative evidence matters: Unlike traditional methods, DhondtXAI distinguishes between features that push a prediction in one direction versus another. This separation, combined with optional "feature alliances" (grouping related features), reduced projection error from 0.2527 to 0.0001 in interaction-heavy scenarios, making explanations more accurate and interpretable.

High agreement with existing methods, with added flexibility: DhondtXAI achieved strong correlation with SHAP (Spearman correlation of 0.93 on real healthcare datasets), but offers additional customization through thresholds and alliances. This positions it as a complementary tool rather than a replacement, giving practitioners more options for explaining AI decisions in healthcare and other domains.

Completeness is mathematically guaranteed: The method preserves "completeness"—meaning all of the model's output difference is accounted for in the explanation—by construction. The framework reports a diagnostic metric (projection residual ratio) so users can verify this property holds in their specific application.

Real-world validation on healthcare data: Testing on Wisconsin Diagnostic Breast Cancer and diabetes risk prediction datasets demonstrates practical applicability, suggesting the method could help clinicians and patients better understand AI-assisted medical decisions.