Title: DhondtXAI: A Democratic Approach to Explainable AI for Feature Attribution in D'Hondt Projections

Abstract: This paper introduces DhondtXAI, a novel attribution framework for tabular Explainable AI (XAI) that operates independently of SHAP, utilizing the D'Hondt method instead. Rather than relying on model-native feature importance metrics or SHAP values, DhondtXAI calculates background-interventional removal effects. The framework distinguishes between positive and negative evidence, allows for the formation of optional feature alliances, and incorporates optional thresholds. It then allocates "seats" according to the D'Hondt rule and projects these results onto the local difference in model output. By design, the method ensures completeness, while the projection residual ratio is provided as a diagnostic metric.

The approach was subjected to a comprehensive evaluation involving synthetic tests for additive and interaction effects, perturbations of correlated features, ablations regarding operators and apportionment, comparisons of projection modes, checks at the logit scale, repeated split validations, paired deletion tests, and applications to two healthcare datasets. These datasets include the Wisconsin Diagnostic Breast Cancer dataset, modeled using CatBoost, and a dataset for early-stage diabetes risk prediction, modeled using XGBoost. SHAP was utilized solely as an external benchmark with aligned parameters.

In synthetic additive scenarios, DhondtXAI precisely recovered the ground-truth rankings. For multiplicative interactions, the use of alliances decreased the mean projection residual from 0.2527 to 0.0001. On the WDBC and diabetes datasets, the method demonstrated strong agreement with SHAP, yielding Spearman correlation coefficients of 0.9273 and 0.9353, respectively. These findings were further corroborated by additional analyses focusing on signed values, top-k features, magnitudes, deletions, and sensitivity. The results suggest that DhondtXAI serves as a complementary tabular XAI tool—characterized by its proportional, alliance-aware, and threshold-aware nature—rather than a direct replacement for existing methods like SHAP or LIME.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC