Model-agnostic Ensemble-based Explanation Correction Leveraging Rashomon Effect
Masaki Hamamoto, Masashi Egi · 2021 IEEE Symposium Series on Computational Intelligence (SSCI) · 2021
Explainable artificial intelligence (AI) technology enables us to quantitatively analyze the whole prediction logic of AI as a global explanation. However, unwanted relationships learned by AI due to data sparsity, high dimensionality, and noise are also visualized in the explanation, which deteriorates confidence in the AI. Thus, establishing a methodology of correcting those unwanted relationships in an explanation has been developed. We propose a model-agnostic ensemble-based explanation correction (EBEC) leveraging the Rashomon effect of statistics and evaluated it. The evaluation results indicate that the EBEC can correct a targeted part of the global explanation of AI so that the part aligns with the domain knowledge given by the user while maintaining other parts of the explanation. The results also indicate that the EBEC can improve the prediction accuracy of AI when appropriate domain knowledge was given.