SAGE-GCN: Graph Convolutional Network Based on Self-adaptive Stable Gates for Link Prediction in Dynamic Complex Networks

Abstract

<p>Link prediction is one of the most important tasks in uncovering evolving mechanisms of dynamic complex networks. Existing dynamic link prediction models suffer from limitations such as vulnerability to adversarial attacks, poor accuracy, and instability. In this paper, we propose a novel dynamic Graph Convolutional Network model incorporating a Self-adaptive Stable Gate (SAGE-GCN) consisting of a state encoding network and a policy network. Firstly, we capture the local topology of the nodes by employing a multi-power adjacency matrix to obtain higher-order topological features, enabling its features to be distinguished at different network snapshots. Then, a stable gate is introduced to ensure multiple spatiotemporal dependency paths within the state encoding network. It is proven that SAGE-GCN with integral Lipschitz graph convolution is stable to relative perturbations in the dynamic networks. Finally, a self-adaptive strategy is proposed to choose different state encoding network instances, with a policy network used to learn the optimal temporal and structural features through corresponding rewards to capture network dynamics. With the aid of extensive experiments on five real-world graph benchmarks, SAGE-GCN is shown to substantially outperform current state-of-the-art approaches in terms of precision and stability of dynamic link prediction and ability to successfully defend against various attacks.</p>