Performance-Interpretability Tradeoff of Mamdani Neuro-Fuzzy Classifiers For Medical Data

Abstract. Neuro-fuzzy systems are models that incorporate the learning abil-ity and performance of Artificial Neural Networks (ANNs) with the inter-pretable reasoning of fuzzy inference systems (FISs). An ANN can learn pat-terns from data and achieve high accuracy, while a FIS uses linguistic and interpretable rules to match inputs and outputs of the data. Two types of FISs are used the most in literature: Takagi-Sugeno-Kang (TSK) and Mamdani. The main focus of this paper is on the Mamdani neuro-fuzzy sys-tems, notably the Hybrid Neuro-Fuzzy Inference System (HyFIS) and the Neuro-Fuzzy Classifier (NEFCLASS). It aims at evaluating and comparing the two classifiers over two medical datasets in order to determine which one can offer the best performance-interpretability tradeoff. Results show that HyFIS is the best in terms of performance, while NEFCLASS is better in terms of interpretability. As for the performance-interpretability tradeoff, NEFCLASS is the best; it achieves a good performance while being less complicated and more interpretable.