Hyper-Ordinal Pattern: Measuring High-Order Connection Relationship in Brain Disease Networks

Brain hyper-networks as a kind of hypergraph for brain network analysis, describing the high-order interactions among brain regions, have been extensively utilized in research on brain diseases such as mild cognitive impairment (MCI) and Alzheimer's disease (AD). Recently, several hypergraph representation methods (e.g., hypergraph neural network) have been constructed for investigating brain hyper-networks. However, most of the existing methods neglected the connection relationships on hyperedges with important weighted information in brain hyper-networks. To tackle this problem, we propose hyper-ordinal pattern (HOP) as a novel approach for representing and analyzing brain hyper-networks. Different from the existing hypergraph representation studies, we utilize the ordinal pattern relationships on hyperedges with weighted information to construct HOPs for brain hyper-networks. In HOP, each node is concatenated via ordinal hyperedges, where hyperedge weights have ordinal pattern relationships (e.g., descending order). We construct HOP for each node in brain hyper-networks and propose node HOP (NHOP) kernel for measuring node similarity in brain hyper-networks. Based on NHOP kernel, we further present a new brain hyper-network kernel called ordinal pattern based hyper-network (OPHN) kernel to calculate the brain hyper-network similarity. In order to assess the performance of the proposed OPHN kernel, we conduct the extensive experiments in brain hyper-networks of brain diseases including MCI and AD. The experimental results demonstrate that the proposed OPHN kernel is better than the state-of-the-art competing methods (e.g., hyper-graph neural network) in the classification tasks of brain diseases. Additionally, the proposed NHOP kernel can effectively identify altered hyper-ordinal patterns in the brain hyper-networks of patients with brain diseases.