Analysis And Research On The Quasi-linear SVM Training Subsets Divided

This paper proposes an improved hierarchical clustering method based on PCA-clusters merging for quasi-linear SVM in order to fulfill classification tasks. The quasi-linear SVM is a SVM with quasi-linear kernel function. The quasi-linear kernel realizes that a nonlinear separating boundary between class labels is approximated by a series of local linear boundaries.1) Separating Boundary DetectionIn some real world dataset, there are some tasks with noise. A guided partitioning approach of separating boundary detection is proposed to partition training data into subsets. The separating boundary detection algorithm considering sample label changes in neighbor area is used to select samples which near the nonlinear separating boundary.The selected samples are partitioned into local subsets, the prior knowledge of local subsets is used capture the distribution information of separating boundary. To some extent, this algorithm can improve the accuracy of quasi-linear classification.2) Hierarchical ClusteringThe hierarchical clustering method is used for merging the closest pair of objects in one local linear boundary, which reduces the number of the clusters under a specific value.3) Cluster Merging based on PCAAs the subset captured by hierarchical clustering method can not have local linear property. The merging operation of the clusters which are distributed around the local linear boundaries, has been implemented according to the angular relationship between the PCA vectors of the neighboring clusters.The prior knowledge of local clusters has been used to construct a composite kernel, then the quasi-linear SVM is realized by using the composite kernel exactly in the same way as a conventional SVM model. Experimental results on benchmark datasets demonstrate that the proposed method improves the classification performance efficiently.