Endorsed Attributions: eXplainable AI (XAI) with Voting Mechanism with Application in Healthcare

Feature attribution methods have been widely employed in the eXplainable Artificial Intelligence (XAI) community and they have been designed with various underlying concepts. As a consequence, these methods sometimes assign different feature importance on the same data. To resolve the possible discrepancies, we introduce \textit{endorsed attributions}, a voting mechanism that integrates the strengths of different feature attribution methods on healthcare datasets: maternal health risk classification, body signal of smoking, fetal health and heart disease classification. Our methods generate simplified visualization charts as an explainable AI use case in clinical settings, avoiding the need for end users like patients or medical practitioners to focus too deeply on technical nuances. Graph visualizations from our partitioning methods also reveal the differences between correct and wrong predictions. Finally, we demonstrate how the extraction of endorsement cores (EEC) can be used for data pruning, and models trained on the resulting EEC data subset can achieve competitive performance.