Performance Analysis of Granular versus Traditional Neural Network Classifiers: Preliminary Results

Abstract

A recent trend in Machine Learning is to augment the transparency of traditional classification models using Granular Computing techniques. This approach has been found particularly useful in the neural networks field since most successful neural systems often require complex structures to behave like universal approximators. However, there is a widely-held view stating that, to build an interpretable classifier, one might have to sacrifice some prediction accuracy. We want to challenge this belief by exploring the performance of a recently introduced granular classifier termed Fuzzy-Rough Cognitive Networks against low-level (i.e., traditional) neural networks. The simulation results have shown that this neural system can attain quite competitive prediction rates while featuring a shallow, granular architecture. As a bigger picture, this study paves the way for conducting a more thorough evaluation of granular versus low-level neural classifiers in the near future.