SVM Fuzzy Hierarchical Classification Method for Multi-class Problems

In this paper we present a new fuzzy classification method based on support vector machine (SVM) to treat multi-class problems. Generally, SVMs classifiers are designed to solve binary classification problem. In order to handle multi-class classification problem, we present a new method to build dynamically a fuzzy hierarchical structure from the training data. Our method is based on two main concepts: fuzzy hierarchical classification and support vector machine. First, the fuzzy hierarchical classification consists in finding relationships between objects. We introduce the transitive closure measure to discover fuzzy similarity between objects. Second, SVM is applied at each node of the hierarchy to discriminate between objects. SVM is used to divide the original problem into sub-problems. We combine multiple binary SVMs to solve multi-class classification. We use equivalence classes to regroup similar objects into single class. Finally, we get a direct hierarchy of classes. Our experimental results show that the proposed model of fuzzy classification is very effective and efficient to handle multiclass problem.