Time-constrained walk fused memory enhanced temporal graph neural network

Abstract:

To address the challenges faced by existing temporal graph representation learning methods in fully exploiting structural features and dependencies of the network, a time-constrained walk fused memory enhanced temporal graph neural network (TWMTGN) is proposed. Firstly, a specific type of temporal constraint walk sequence is constructed for the interaction node, and the memory module is used to capture the long-term dependence of the network. The walk sequence features are fused into the memory of interaction node, which is dynamically updated upon the occurance of events. Secondly, a feature attenuation layer is designed according to the node type and time interval to model short-term dependencies, which improves the model′s ability to identify key historical interaction nodes. Finally, the aggregated historical interaction features of the target node are fed into a causal convolutional network to further uncover the potential association between them. Experimental results on real datasets show that the proposed network can improve the performance of temporal link prediction tasks while maintaining relatively low complexity. Parameters such as the length and frequency of time-constrained walks affect the mode′s performance. The proposed time-constrained walk sequences can effectively capture the structural features of the network, while the node memory and feature attenuation layers help to capture the network dependencies.