SE-GRU: Structure Embedded Gated Recurrent Unit Neural Networks for Temporal Link Prediction

Temporal link prediction on dynamic graphs is essential to various areas such as recommendation systems, social networks, and citation analysis, and thus attracts great attention in both research and industry fields. For complex graphs in real-world applications, although recent temporal link prediction methods perform well in predicting high-frequency and nearby connections, it becomes more challenging when considering low-frequency and earlier connections. In this work, we introduce a novel and elegant prediction architecture called Structure Embedded Gated Recurrent Unit (SE-GRU) neural networks, to strengthen the prediction robustness against frequency variation and occurrence delay of connections. The established SE-GRU embeds the structure for local topological characteristics to emphasize the different connection frequencies between nodes and captures the temporal dependencies to avoid losing valuable information caused by long-term changes. We realize neural network optimization considering three terms concerning reconstruction, structure, and evolution. The extensive experiments performed on three public datasets demonstrate the significant superiority of SE-GRU compared with 5 representative and state-of-the-art competitors under three evaluation metrics. The results validate the effectiveness and robustness of our proposed method, by showing that the frequencies and timestamps of connections have a little-to-no negative impact on prediction accuracy.

Automated summary

This article introduces a novel prediction architecture called SE-GRU neural networks for temporal link prediction on dynamic graphs. By embedding structure and capturing temporal dependencies, this method effectively handles frequency variation and occurrence delay, resulting in robust predictions.