Structural and topological guided GCN for link prediction in temporal networks

The ever-growing size of social network information has enhanced research aimed at finding solutions to challenges in this arena. The vastness and complexity of interactions between social network entities render link prediction in these datasets a challenging task. Previous studies often concentrate on only exploring the local node connectivity information neglecting other key network-characterizing properties. In addition, most works assume static networks, yet many real-world graphs evolve. To address these limitations, firstly, we explore topological information from input graph adjacency matrices by computing topological similarity-based convolution feature matrices. Secondly, we leverage the node strength centrality matrix, a more powerful variant of node degree to preserve the node centrality roles and node’s structural connectivity information throughout the network. Lastly, we deploy an LSTM layer to explore the underlying network temporal information. The proposed Structural and Topological aware GCN (STP-GCN) is tested on five social network datasets. Based on experimental results, it exhibits a 3% link prediction AUC improvement, negligible training time increment per epoch (0.2s), and a large MSE magnitude (2.5) reduction in structural centrality prediction as compared to the best benchmark.