M-LIME: a marginalized local interpretable model-agnostic explanations

To mitigate the inherent black-box nature of deep models, explainable artificial intelligence has gradually received widespread attention. Local Interpre table Model-agnostic Explanations is a popular interpretable algorithm, but its random deletion of perturbed samples generation leads to the problems of unstable interpretation results and out-of-distribution of the perturbed samples, resulting in a significant reduction in the credibility of the interpretation results. To solve the above problems, this paper proposes an explainable algorithm for text classification models based on input marginalization, alleviating its many problems through external component enhancement. The experimental results show that the method proposed in this paper has improved stability and effectiveness with low time cost. In addition, an evaluation metric is proposed that can more comprehensively and accurately measure the interpretable algorithm.