Prototype-Optimized FCM: A Framework for Interpretable Fuzzy Classifier Synthesis

Fuzzy C-Means (FCM) and its variants are prominent unsupervised clustering techniques, renowned for their robustness in handling data ambiguity. Such methods can also be applied to the design of supervised classifiers by involving the participation of real labels in the optimization of the algorithm. This study develops a novel prototype optimization strategy for enhancing the performance of the fuzzy classification. The developed strategy systematically refines the initial cluster prototypes and concurrently optimizes the associated fuzzy coefficients, enabling the learning of data-specific prototype configurations that directly maximize class separability. By iteratively adjusting both the prototypes and the fuzzy membership coefficients, the algorithm achieves a more discriminative feature space where data points from different classes are better separated. This refinement process is guided by a supervised perturbation mechanism that balances the compactness of clusters with the separation between them, ensuring robustness against noise and outliers. Furthermore, the proposed strategy leverages the inherent uncertainty in fuzzy clustering to explore a wider range of potential prototype configurations, thereby enhancing the performance of the method to discover optimal class boundaries. The model is rigorously evaluated on several benchmark datasets, and the results show an improvement in classification performance.