Mechanistic Interpretation of Machine Learning Inference: A Fuzzy Feature Importance Fusion Approach

With the widespread use of machine learning to support decision-making, it is\nincreasingly important to verify and understand the reasons why a particular\noutput is produced. Although post-training feature importance approaches assist\nthis interpretation, there is an overall lack of consensus regarding how\nfeature importance should be quantified, making explanations of model\npredictions unreliable. In addition, many of these explanations depend on the\nspecific machine learning approach employed and on the subset of data used when\ncalculating feature importance. A possible solution to improve the reliability\nof explanations is to combine results from multiple feature importance\nquantifiers from different machine learning approaches coupled with\nre-sampling. Current state-of-the-art ensemble feature importance fusion uses\ncrisp techniques to fuse results from different approaches. There is, however,\nsignificant loss of information as these approaches are not context-aware and\nreduce several quantifiers to a single crisp output. More importantly, their\nrepresentation of 'importance' as coefficients is misleading and\nincomprehensible to end-users and decision makers. Here we show how the use of\nfuzzy data fusion methods can overcome some of the important limitations of\ncrisp fusion methods.\n