Abstract: Abstract: As a key technique in network evolution analysis, link prediction aims to infer potential future connections based on historical interaction data. Its accuracy directly impacts the effectiveness of critical applications such as personalized recommendation and disease spread prediction. However, existing link prediction algorithms still face three major limitations when applied to dynamic networks: (1) missing node attributes result in insufficient feature representation capacity; (2) the spatiotemporal dynamics of networks make algorithm design challenging; and (3) data sparsity leads to inadequate model training. These issues significantly hinder the practical application of current link prediction methods in real-world dynamic network scenarios. To address the above challenges, this paper proposes a node encoding-based generative adversarial dynamic link prediction algorithm, named Attention-GAN. First, a node feature encoding module is constructed using positional encoding and random walk techniques to generate unique initial representations for each node, effectively mitigating the problem of missing attributes. Second, a spatial feature extraction module based on graph convolutional networks (GCNs) is employed to perform graph convolution operations on the feature matrix of each snapshot, capturing local structural information. Third, a temporal feature extraction module based on a multi-head attention mechanism is designed to capture temporal dependencies from multiple perspectives in parallel. This module incorporates residual connections and layer normalization to enhance the training stability of deep models. Finally, a generative adversarial training framework is constructed to alleviate underfitting caused by data sparsity. Experimental comparisons with five state-of-the-art link prediction algorithms on four public datasets demonstrate the effectiveness and superiority of the proposed Attention-GAN approach. Experimental results show that Attention-GAN achieves performance improvements of approximately 7.57%, 22.22%, and 13.41% in AUC, MR, and Recall, respectively, compared to the sub-optimal algorithm.

Key words: Link prediction, deep learning, generative adversarial network, graph convolutional network, attention mechanism

摘要： 作为网络演化分析的关键技术，链路预测旨在通过历史交互数据推断未来潜在连接关系，其准确性直接影响个性化推荐、疾病传播预测等重要应用场景的效能。然而，现有链路预测算法在面对动态网络时仍存在三个明显局限：（1）节点属性缺失导致特征表达能力不足；（2）网络时空动态性导致算法难以设计；以及（3）数据稀疏性引发的模型训练不充分。这些限制严重制约了现有链路预测算法在针对动态网络真实场景的应用效果。为有效处理上述问题，本文提出了一种基于节点编码的生成对抗式动态链路预测算法Attention-GAN。首先，采用位置编码与随机游走技术构建节点特征编码模块，为每个节点生成唯一的初始表示，有效缓解节点属性缺失问题；其次，利用图卷积神经网络构建空间特征提取模块，对每个图快照的特征矩阵执行图卷积操作，以提取局部结构信息；接着，设计基于多头注意力机制的时序特征提取模块，通过多头机制并行捕获不同视角下的时序依赖关系，并引入残差连接与层归一化以增强深层模型的训练稳定性；最后，构建基于生成对抗网络的对抗训练框架来缓解因数据稀疏性带来的训练不足问题。在5个公开数据集上与5种现有链路预测算法的实验对比证实了所提算法Attention-GAN的有效性和优越性，实验结果表明，相较于次优算法，Attention-GAN在AUC、MR和Recall指标上分别获得了7.57%、22.22%和13.41%左右的性能提升。

关键词: 链路预测, 深度学习, 生成对抗网络, 图卷积网络, 注意力机制