All-Pairs Evolving Fuzzy Classifiers for On-line Multi-Class Classification Problems

Abstract

In this paper, we propose a novel design of evolving fuzzy classifiers in case of multi-class classification problems. Therefore, we exploit the concept of all-pairs aka all-versus-all classification using binary classifiers for each pair of classes, which has some advantages over direct multi-class as well as one-versus-rest classification variants. Regressionbased as well as singleton class label fuzzy classifiers are used as architectures for the binary classifiers, which are evolved and incrementally trained based on the concepts included in the FLEXFIS family (a connection of eVQ and recursive fuzzily weighted least squares). The classification phase considers the preference levels of each pair of classes stored in a preference relation matrix and uses a weighted voting scheme of preference levels, including reliability aspects. The advantage of the new evolving fuzzy classifier concept over single model (using direct multi-class classification concept) and multi model (using one-versus-rest classification concept) architectures will be underlined by empirical evaluations and comparisons at the end of the paper based on high-dimensional real-world multi-class classification problems.