Hierarchical support vector machines

The speed and accuracy of a hierarchical SVM (H-SVM) depend on its tree structure. To achieve high performance, more separable classes should be separated at the upper nodes of a decision tree. Because SVM separates classes at feature space determined by the kernel function, separability in feature space should be considered. In this paper, a separability measure in feature space based on support vector data description is proposed. Based on this measure, we present two kinds of H-SVM, binary tree SVM and k-tree SVM, the decision trees of which are constructed with two bottom-up agglomerative clustering algorithms respectively. Results of experimentation with remotely sensed data validate the effectiveness of the two proposed H-SVM.