Classification of Directed Networks With Application to Neuroimaging Data

Graph-based classification techniques applied to neuroimaging data have played a pivotal role in advancing our comprehension of brain functionalities. Existing graph-based classification methods primarily focus on undirected networks. However, networks derived from neuroimaging data often exhibit a directed nature. In this study, we propose an innovative asymmetric bilinear logistic regression (ABLR) approach to tackle binary classification tasks in weighted directed networks. This general framework is achieved by incorporating the directional flow information of edge sending and receiving. Simultaneously, the loss function is equipped with regularization penalties. This way, our method can identify predictive nodes sparsely and generate meaningful interpretations. To solve the optimization problem, we develop an efficient proximal linear block coordinate descent (prox-linear BCD) algorithm, which is proved to have a global convergence property. Through simulations and electroencephalogram(EEG)-derived brain network application, our proposed classification method outperforms the alternative method. The ABLR method not only achieves higher classification accuracy but also provides stronger interpretability.