A Novel Algorithm For Multi-class Classification Based On Multi-space Mapped Support Vector Machine

The multi-classification problem always arouses reseachers’attention in pattern recognition and machine learning, which is still hot spot solved with more effective methods although binary classification for support vector machines(SVM) has been applied in the problem successfully. Considering the time-consuming weakness of all-together and decomposed methods cause more classifiers and longer testing time, the hybrid approach uses binary tree as base structure for partitioning datasets and SVM to train the base learner in nonleaf node became one of the most wathched method. Owing to the mean of node’s datasets division has a direct influence towards the performance of this algorithm, so it is extremely important.Liu used progressive K-means algorithm to partition the data and proposed multi-mapped SVM which acquired better performance, however, the measure is so random that the performance is not steady. Consequently, a new technique based on multi-space method with above hybrid structure is presented in this paper. The method first defines a classifier score function, simple and easy to compute, to find the best data-division for every nonleaf node. Furthermore, for the sake of "combination explosion" caused by application to much larger classes, special Genetic Algorithm(GA) is added to contract searching space and find the best solution.To compare the performance with different measures for data-division, we use K-means, kernel clustering distance measure and classifier score function to conduct a test. And the testing results of eight UCI datasets and two face recognition datasets show that our proposed algorithm, using classifier score function, is better than the former two measures, not just in generalization ability but also on the stabilization. Moreover, our classifier score function is easy to compute, so the time complexity is the lowest in three measures through the comparison of the create-tree time, and the algorithm shows more superiority especially in the datasets with larger classes.