Robust spatio-temporal graph neural networks with sparse structure learning

Abstract

This paper focuses on the problem of spatio-temporal graph classification by introducing sparse structure learning to enhance its robustness and explainability. Spatio-temporal graph neural networks (STGNN) integrate spatial structure and temporal sequential features into GNN learning, resulting in promising performance in many applications. However, current STGNN models often fail to capture the discriminative sparse substructure and the smooth distribution of these samples. To this end, this paper introduces RostGNN, robust spatio-temporal graph neural networks, for achieving more discriminative graph representations. Concretely, RostGNN extracts the spatial and temporal features by performing gated recurrent units on the given time series data and calculating adjacent matrixes for graphs. Then, we impose the iterative hard-thresholding approach on the final association matrix to obtain a sparse graph. Meanwhile, we calculate a similarity matrix from the side information of samples to smooth the achieved data representations and use fully connected networks for graph classification. We finally applied RostGNN to brain graph classification in experiments on real-world datasets. The results demonstrate that RostGNN delivers robust and discriminative graph representations and performs better than compared methods, benefiting from the sparsity and manifold regularizers. Furthermore, RostGNN can potentially yield useful findings for data understanding.