An asymmetric graph neural network with time-position encoding for link prediction in dynamic networks

Abstract In dynamic networks, the network structure continuously changes over time, such as in social networks and transportation systems. Accurately predicting links in dynamic networks is an extremely difficult task. Although existing Temporal Graph Neural Networks (TGNNs) have demonstrated considerable success in predicting links in dynamic networks, their symmetric evaluation strategy limits their scalability in dynamic scenarios. For example, in full-node evaluation, these symmetric update methods not only require simultaneous updates of both source and target nodes but also trigger cascading updates in the neighborhood context, resulting in significant computational overhead. To control this cost, existing TGNNs models have to compromise between temporal proximity and topological structure: either truncate historical context leading to the loss of long-term dependencies, or simplify neighborhood sampling resulting in over-smoothing. This trade-off severely limits model performance.To address the above challenges, this paper propose the TPGNN model, which is an asymmetric temporal graph neural network that only updates source nodes. During evaluation, by updating only the source nodes and precomputing and freezing the target nodes and candidates, it eliminates cascading updates while also improving performance. Additionally, the model introduces inter-layer smoothing, residual connections, multi-scale spatio-temporal encoding, and position-aware attention modules. In link prediction experiments, TPGNN achieves significant performance improvements on four public datasets, while maintaining CPU-friendly efficiency, enabling real-time evaluation in resource-constrained scenarios. The code has been released at: https://github.com/LiMinlr/TPGNN .