基于双层时序的动态社交网络链路预测

Dynamic Social Network Link Prediction Based on Bi-Layer Temporal Modeling

摘要: 现实世界中存在大量网络图信息，其中社交网络实体间交互的动态性和复杂性使得动态图链路预测成为了一项更具有挑战性的任务。传统基于图神经网络的动态链路方法由于过平滑性往往只关注图中局部特征，难以获取图中实体的全面性信息，并且发现在社交网络中的连通接近性对未来链路预测是有利的。为了解决以上挑战，本文设计了双层时序模型Bi-GTGNN。首先提取每个快照的子图，并将每个快照的子图集抽象为时序序列，然后设计全局时序图神经网络提取图的全局信息并生成快照表示。其次，将每个时间戳的快照表示输入到LSTM中进一步提取时序信息，并设计了新颖的损失函数训练具有连通接近性的图嵌入。最后将具有时序信息的图嵌入用于链路预测。在五个数据集上进行了大量实验，结果表示Bi-GTGNN性能优于其它先进的baseline模型。

Abstract: In the real world, there is a plethora of network graph information, where the dynamism and complexity of interactions among entities in social networks have made dynamic graph link prediction a more challenging task. Traditional methods based on graph neural networks for dynamic links often focus only on local features of the graph due to over-smoothing, making it difficult to acquire comprehensive information about entities in the graph. Additionally, we observe that the connectivity proximity in social networks is advantageous for future link prediction. To address these challenges, we propose a bi-layer temporal model. Firstly, we extract subgraphs for each snapshot and abstract the subgraph set of each snapshot into a temporal sequence. Then, we design a global temporal graph neural network to extract the global information of the graph and generate snapshot representations. Secondly, we input the snapshot representations of each timestamp into an LSTM to further extract temporal information and design a novel loss function to train graph embeddings with connectivity proximity. Finally, the graph embeddings with temporal information are utilized for link prediction. We conduct extensive experiments on five social network datasets, and the results demonstrate that our model outperforms other state-of-the-art baseline models.