A clustering based adaptive DAG for multiclass Support Vector Machine

Abstract

This paper presents a method for multiclass Support Vector Machine(MCSVM), which we called CLustering Adaptive Directed Acyclic Graph(CLADAG). A previous approach, the Decision Directed Acyclic Graph(DDAG) is proposed to half randomly select a classifier from a set of classifier which is produced in the training phase. Using DDAG, the testing result of the unlabeled sample may be different if the label of some classes is swapped, leading to a unstable classification accuracy. In order to get definite testing result for the same sample, we use a heuristic method based on clustering to sort the order of classifier for all unlabeled samples. The experimental results demonstrated CLADAG is an effective method with definite results.