Answers are grounded in this paper and cite their sources.
Temporal Link Prediction (TLP) is a key task in social network research, aiming to predict future network links based on the relative behaviors of previous network changes. The various interaction patterns and nonlinear characteristics of the dynamic network pose challenges for the design of accurate general prediction algorithms. With the development of deep learning technology, graph neural networks (GNNs) have been widely applied to solve graph-related problems and have achieved significant results. However, solving the TLP problem based on GNNs usually involves complex memory update mechanisms, multi-layer iterative message passing paradigms, and reliance on negative sampling strategies, all of which contribute to the overall model’s complexity. To reduce model complexity while maintaining high accuracy, this paper proposes a lightweight temporal aggregation graph neural network model (LAGNN) that uses only a single layer to solve the TLP problem. This model decouples the static identity representation of nodes from their dynamic interaction context, simplifying the message passing mechanism and training objective while preserving the accuracy of temporal link predictions and significantly improving efficiency. The core of LAGNN lies in decoupling representation modeling: static embeddings capture inherent node identities, and a temporal attention module aggregates historical neighbor information to generate context‑aware dynamic node representations. During training, LAGNN optimizes an objective function that maximizes the similarity between positive samples, thereby avoiding the inherent instability and bias of negative sampling strategies. At inference time, the model uses the static embeddings of all candidate nodes and only needs to compute the dynamic representation of the source node. Experimental results on temporal link prediction tasks using four public dynamic social network datasets demonstrate that, despite using only a single‑layer architecture, LAGNN achieves predictive accuracy comparable to or even surpassing state-of-the-art multi-layer methods while also validating its efficiency and stability advantages.
No takes yet. Share an insight, caveat, or question.
Han et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: