FLARE: Fuzzy Local Agnostic Rule-Based Explanations for Closed Box Classifiers

The massive amount of data available in recent years has led to explosive growth in the field of machine learning. <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Closed box</i> models gain from this because they can be trained more effectively and provide better results than <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">Glass box</i> models, albeit at the cost of interpretability. Critical applications such as medicine or aviation would benefit greatly from these more powerful models, but they cannot afford to trust a closed box model unquestioningly; hence, they need a way to explain these methods (or at least their reasoning when classifying an instance). A well-accepted practice is to build a surrogate <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">explainable</i> (Glass box) model that represents a small region around the instance to be explained by extracting factual and counterfactual explanations. A factual explanation justifies the classification of the given instance into a specific category, whereas a counterfactual explanation points out why that particular instance has <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">not</i> been classified differently. This work introduces FLARE, an agnostic algorithm that employs fuzzy logic to generate factual and counterfactual explanations. First, it generates a synthetic neighborhood around the instance. Then, it learns a fuzzy decision tree using the neighborhood. After that, factual and counterfactual explanations are extracted from the tree. Finally, to maintain the semantics of the problem domain, the explanations are mapped to the global fuzzy sets, defined over the entire range of each variable. Experiments have been conducted using well-known closed box algorithms, datasets, and state-of-the-art counterfactual methods. The results show that FLARE performs as well as the current best methods while being computed orders of magnitude faster.