A hierarchical Clustering Method Based on PCA-Clusters Merging for Quasi-linear SVM

This paper proposes an improved hierarchical clustering method based on PCA-clusters merging for quasi-linear SVM. The quasi-linear SVM is an SVM with quasi-linear kernel. It considers a nonlinear separating boundary between class labels as an approximation of multiple local linear boundaries with interpolation and the quasi-linear kernel is composited based the information of local clusters along the boundary. In order to obtain the local clusters, the proposed clustering method, first detects the nonlinear boundary based on the changes of class labels; then obtains small partitions along the nonlinear separating boundary using a hierarchical clustering; and further merges the nearest neighboring clusters distributed in one local linear boundary into one cluster according clusters distributed in one local linear boundary according to PCA-based criterion. The quasi-linear kernel is composited based on the information of local clusters. Experimental results on benchmark datasets demonstrate that the proposed method improves the classification performance efficiently.