Threshold-attention graph neural network for dynamic link prediction

Accurately predicting future connections between nodes in dynamic networks is crucial for uncovering network evolution patterns and enabling practical applications, where Spatiotemporal Graph Neural Network (ST-GNN)-based methods have rapidly emerged as a promising technique. Despite the recent advances in ST-GNNs, two fundamental challenges persist from a scientific perspective: (1) the efficient integration of globally relevant structural information while mitigating noise from irrelevant or redundant features remains an open problem in graph representation learning; (2) capturing intricate multi-scale temporal dependencies in time-varying networks poses significant difficulties, as conventional RNNs often fail to model multi-scale dynamic patterns effectively. To address these challenges, this paper proposes a novel threshold-attention graph neural network (TAGNN). First, at the spatial global feature extraction level, we designed a threshold attention that adaptively generates thresholds for each node-pair based on node features and attention scores to exclude nodes with score below the threshold from the aggregation process, thereby reducing redundant information. Second, for temporal modeling, we innovatively introduce a U-Net Fusion Gated Unit (UFGU), which leverages U-Net’s multi-scale architecture to capture complex temporal correlations that traditional RNNs often miss. Extensive experiments on four real-world dynamic network datasets demonstrate that TAGNN achieves significant performance gains, such as AUC improvements of 1.05% and 2.02% on the UCI and Dept datasets, respectively, outperforming state-of-the-art baselines. Ablation studies confirm the individual and synergistic contributions of the proposed threshold-attention and UFGU modules.