Signal Subgraph Estimation Via Vertex Screening

Graph classification and regression have wide applications in a variety of domains. A graph is a complex and high-dimensional object, which poses great challenges to traditional machine learning algorithms. Accurately and efficiently locating a small signal subgraph dependent on the label of interest can dramatically improve the performance of subsequent statistical inference. Moreover, estimating a signal subgraph can aid humans with interpreting these results. We present a vertex screening method to identify the signal subgraph when given multiple graphs and associated labels. The method utilizes distance-based correlation to screen the vertices, and allows the subsequent classification and regression to be performed on a small induced subgraph. We demonstrate that this method is consistent in recovering signal vertices and leads to better classification performance via theory and numerical experiments. We apply the vertex screening algorithm on human and murine graphs derived from functional and structural magnetic resonance images to analyze the site effects and sex differences.