GridEx: An Algorithm for Knowledge Extraction from Black-Box Regressors

Knowledge-extraction methods are applied to ML-based predictors attain explainable representations of their operation when the lack interpretable results constitutes a problem. Several algorithms have been proposed for knowledge extraction, mostly focusing on extraction either lists or trees rules. Yet, most them only support supervised learning – and, in particular, classification tasks. Iter is among few rule-extraction capable extracting symbolic rules out sub-symbolic regressors. However, its performance here intended as interpretability it extracts easily degrades complexity regression task at hand increases. In this paper we propose GridEx, an extension algorithm, aimed form if-then-else from any sort regressor—there including neural networks arbitrary depth. With respect Iter, GridEx produces shorter rule retaining higher fidelity w.r.t. original regressor. We report several experiments assessing against and Cart (i.e., decision-tree regressors) used benchmarks.