Evaluating Explainable Artificial Intelligence Methods Based on Feature Elimination: A Functionality-Grounded Approach

Abstract

Although predictions based on machine learning are reaching unprecedented levels of accuracy, understanding the underlying mechanisms of a machine learning model is far from trivial. Therefore, explaining machine learning outcomes is gaining more interest with an increasing need to understand, trust, justify, and improve both the predictions and the prediction process. This, in turn, necessitates providing mechanisms to evaluate explainability methods as well as to measure their ability to fulfill their designated tasks. In this paper, we introduce a technique to extract the most important features from a data perspective. We propose metrics to quantify the ability of an explainability method to convey and communicate the underlying concepts available in the data. Furthermore, we evaluate the ability of an eXplainable Artificial Intelligence (XAI) method to reason about the reliance of a Machine Learning (ML) model on the extracted features. Through experiments, we further, prove that our approach enables differentiating explainability methods independent of the underlying experimental settings. The proposed metrics can be used to functionally evaluate the extent to which an explainability method is able to extract the patterns discovered by a machine learning model. Our approach provides a means to quantitatively differentiate global explainability methods in order to deepen user trust not only in the predictions generated but also in their explanations.