Hierarchical Representation Learning of Dynamic Brain Networks for Schizophrenia Diagnosis.

The dynamic brain network (DBN) consists of a set of time-varying states (i.e., functional connectivity matrixes), and has revolutionized the field of brain network analysis through captured dynamic evolution patterns. However, current representation methods of DBN, which use hand-crafted features, are still not data-driven, and cannot effectively learn the hierarchically organized temporal nature of DBN. To address this issue, we propose a novel hierarchical representation learning (HARL) method for dynamic brain networks in the framework of graph convolutional networks. Specifically, we first define a graph model, whose nodes represent time-varying states, node features are determined by states’ functional connectivity matrix, and edges learn according to node features. Then, based on this model, we build a HARL module to learn the representation of DBN under different levels. Each level consists of a convolution layer and a pooling layer. Through a convolution layer, node features can be updated according to their neighbor nodes, which can better learn the representation of each state. In the pooling layer, according to both node features and graph topology, we select some important nodes (i.e., states) from the whole graph to form a coarsened graph, which would be further input to the next level. In each level, the representation of DBN is generated by aggregating these node features. These representations will be input to the fully connected layer for disease prediction. Experiments on a real schizophrenia dataset demonstrate the effectiveness and advantages of our proposed method.