Ordinal Pattern Kernel for Brain Connectivity Network Classification
Brain connectivity networks, which characterize the functional or structural interaction of brain regions, has been widely used for brain disease classification. Kernel-based method, such as graph kernel (i.e., kernel defined on graphs), has been proposed for measuring the similarity of brain networks, and yields the promising classification performance. However, most of graph kernels are built on unweighted graph (i.e., network) with edge present or not, and neglecting the valuable weight information of edges in brain connectivity network, with edge weights conveying the strengths of temporal correlation or fiber connection between brain regions. Accordingly, in this paper, we present an ordinal pattern kernel for brain connectivity network classification. Different with existing graph kernels that measures the topological similarity of unweighted graphs, the proposed ordinal pattern kernels calculate the similarity of weighted networks by comparing ordinal patterns from weighted networks. To evaluate the effectiveness of the proposed ordinal kernel, we further develop a depth-first-based ordinal pattern kernel, and perform extensive experiments in a real dataset of brain disease from ADNI database. The results demonstrate that our proposed ordinal pattern kernel can achieve better classification performance compared with state-of-the-art graph kernels.