Beyond the barrier of mistrust – an overview of selected methods for explaining predictions of machine learning models

Abstract

This paper reviews some of the existing explanation techniques aimed to clarify machine learning model predictions. It begins with the recap of terminology and unification of some exiting definitions of loosely defined terms like “interpretability” or “justification”. Next, three clarification methods are presented - LIME, SHAP, and ANCHOR, with an intuitive explanation of how they operate and simplified formalization. In the end, a subjective comparison of the methods is presented in terms of theoretical “explainability” postulates. Conducted analysis indicates that ANCHOR is built on top of the easiest to understand mathematical apparatus, while SHAP possesses the most reliable theoretical foundations. The paper concludes with a discussion on the very correctness of applying such solutions instead of bringing focus to transparent models in the first place