Insights into Interpretability of Neuro-Fuzzy Systems

Neuro-fuzzy networks revealed their proficiency in learning from data, while offering a transparent and somehow interpretable rule-based model. Recent research focused either on the interpretability of the chosen model or on the system performance. Regarding the interpretability, here an index to control the trade-off between complexity and performance, some insights into fuzzy partitions properties, an ideal fuzzy sets shape, and an evaluation of rules are proposed. All the evaluations are made taking into account the required output and performance. A discussion on results of a system built using the Wisconsin Breast Cancer Dataset is performed as a proof of concept.