Dynamic Network Link Prediction Based on Enhanced Graph Convolutional Networks

Dynamic graph representation learning (DGRL) methods based on Graph Convolutional Networks (GCNs) have been the main approaches for dynamic network tasks including link prediction and node classification applied in numerous fields. The current Graph Convolutional Network (GCN)-based methods mostly neglect the differences in importance of neighbor nodes which leads to an unsatisfied information aggregation and hold an excessively computation complexity by employing additional temporal encoders. To address these issues, we propose the Enhanced Graph Convolutional Networks (EGCN) framework for DGRL. First, we define Structurally Enhanced Weighted Resource Allocation (SE-WRA) similarity which introduces a logarithmic enhancement term and a structural interaction term simultaneously to distinguish the importance of neighboring nodes, and further utilizes this similarity to guide the aggregation. Second, to integrate the local and global structural information of dynamic graphs without temporal encoder, we propose a multi-snapshot gradient optimization (MSGO) method based on smoothed second-order moments to update the GCN weight parameters. Specifically, this method defines an exponential moving average (EMA) mechanism with second-order moments, which smooths the gradients at each time step to generate local smoothed gradients for capturing local structural information. Then, we propose an adaptive gradient aggregation strategy that incorporates a dynamic adjustment factor based on global graph structural information to generate a globally adaptive gradient for updating the GCN, thereby obtaining global information. Experiments on six public datasets show the advantage of our EGCN compared with existing baselines, where it has reached the optimum in fourteen out of fifteen performance metrics.