DGSLN: Differentiable graph structure learning neural network for robust graph representations

Abstract

Recently, graph neural networks (GNNs) have been widely used for graph representation learning, where the central idea is to recursively aggregate neighborhood information to update the node feature based on the graph topology. Therefore, an appropriate graph topology is crucial for effective graph representation learning in GNNs. However, most existing GNNs assume that the initial graph is complete and accurate, and utilize the fixed initial graph structure in the entire network, which may limit the learning representation capability of the model. In this work, we propose a novel differentiable graph structure learning neural network (DGSLN), which learns suitable graph structures for GNNs. Specifically, our DGSLN presents a general graph generation scheme that integrates various useful graph prior messages to generate normal structures. We describe the generation process with homophily, node degree, and sparsity as examples. Moreover, we develop a hybrid loss function to ensure the quality of learned graphs, which combines task-specific loss and graph regularization loss to optimize graph structures from both structural adaptive and task-driven aspects. Extensive experiments on graph classification and node classification have shown that our approach significantly improves performance on different benchmark datasets compared to state-of-the-art GNNs methods.