New Graph-Blind Convolutional Network for Brain Connectome Data Analysis

Human connectome provides essential insights in diagnosing many psychiatric disorders. Though machine learning methods in predicting clinical scores have been successfully applied, it is still challenging to capture the complex relation and exploit the graph structure of brain networks. In this paper, we proposed a method to address the problem by extracting the graph embeddings using graph convolutional network (GCN), and using multi-layer perceptron for the regression. Particularly, previous GCN explicitly requires pre-defined graph structures which is not clearly defined in brain connectome. To address this problem, we showed that with naive complete graph structure, GCN can get meaningful results. Meanwhile, an effective algorithm was proposed to learn the graph structure from the data, via generating random graph during training based on the small-world model. Also, the advantages of GCN over multi-layer perceptron was discussed. The experiments demonstrate that the proposed method outperform related baselines significantly on predicting depression scores.