Graph-Kernel Based Structured Feature Selection for Brain Disease Classification Using Functional Connectivity Networks

Abstract

Feature selection has been applied to the analysis of complex structured data, such as functional connectivity networks (FCNs) constructed on resting-state functional magnetic resonance imaging (rs-fMRI), for removing redundant/noisy information. Previous studies usually first extract topological measures (e.g., clustering coefficients) from FCNs as feature vectors, and then perform vector-based algorithms (e.g., $t$ -test) for feature selection. However, due to the use of vector-based representations, these methods simply ignore important local-to-global structural information of connectivity networks, while such structural information could be used as prior knowledge of networks to improve the learning performance. To this end, we propose a graph kernel-based structured feature selection (gk-SFS) method for brain disease classification with connectivity networks. Different from previous studies, our proposed gk-SFS method uses the graph kernel technique to calculate the similarity of networks and thus can explicitly take advantage of the structural information of connectivity networks. Specifically, we first develop a new graph kernel-based Laplacian regularizer in our gk-SFS model to preserve the structural information of connectivity networks. We also employ an $l_{1}$ -norm based sparsity regularizer to select a small number of discriminative features for brain disease analysis (classification). The experimental results on both ADNI and ADHD-200 datasets with rs-fMRI data demonstrate that the proposed gk-SFS method can further improve the classification performance compared with the state-of-the-art methods.