Explainability Improvement Through Commonsense Knowledge Reasoning

Abstract

The explainable artificial intelligence (XAI) studies on the black box model have forcused on a number of area such as feature importance, model agnostic studies, and surrogate model methods. XAI is necessary to increase the explanatory power of feature importance of data and contribute to improving the performance of models. Using human common sense in XAI makes it easier for humans to understand, but such research is lacking. In this paper, we propose a commonsenselearned model and reasoning process to obtain explanatory power which can explain parts that the model could not explain previously in structured data. We extracted common sense about age from ChatGPT, which has recently become a hot topic. Commonsense was used for preprocessing and interpretation of the model results to increase explanatory power and help to understand features. The explanatory power of the model was expressed by Shapley additive explanations and local interpretable model-agnostic explanations and this contributed to the fact that local data could be explained using a commonsense approach learned by humans. abstract environment.