LIPEx-Locally Interpretable Probabilistic Explanations-To Look Beyond The True Class

Hongbo Zhu, Angelo Cangelosi, Procheta Sen, Anirbit Mukherjee

Abstract: The fundamental approach to explanation is to mimic a complex model with a simpler explainer model. Most current work in this field focuses only on explaining the Top predicted class. However, gaining insight into the contributing factors for all potential classes at a particular test point, beyond just the predicted one, can offer valuable insights for machine learning and deep learning practitioners. In this direction, we propose a perturbation-based multi-class explanation model named Locally Interpretable Probabilistic Explanation (LIPEx). LIPEx provides an explanation as a matrix obtained via regression in the space of probability distributions with respect to the squared Hellinger distance. Experiments on both text and image data show that the removal of LIPEx-guided important features from original data causes more prediction degradation of the underlying model than similar tests of other saliency-based or feature importance-based XAI methods. It is also shown that compared to LIME (i.e. state-of-the-art perturbation-based explanation method), LIPEx is more data efficient in terms of using fewer perturbations to obtain a reliable explanation.

External IDs:doi:10.48550/arxiv.2310.04856