Abstract: Real-world networks are often composed of multiple types of entities and interaction relationships, with topological structures and attributes evolving continuously over time. The heterogeneity and dynamics inherent in such networks were fully described by temporal heterogeneous graph. To alleviate the problems of coarse spatiotemporal information fusion and the heavy reliance on manual labels in existing temporal heterogeneous graph representation learning models, this paper proposes a masked autoencoder-enhanced temporal heterogeneous graph representation learning model. First, heterogeneous spatial information was fused through a multi-level attention structure, and temporal information was aggregated across snapshots. Then, representation information of nodes was enriched by leveraging the reconstruction loss of the masked autoencoder. Finally, improvements of 1%–4% in AUC (Area Under the Receiver Operating Characteristic Curve) over baseline models are achieved by the model on link prediction tasks across multiple real-world datasets. An effective self-supervised framework for temporal heterogeneous graph representation learning is provided by the model, facilitating more precise capture of heterogeneous information and dynamic evolution patterns in real networks.

Key words: representation learning, temporal heterogeneous graph, masked autoencoder, self-supervised learning, link prediction

摘要： 现实世界网络往往由多类型实体和交互关系构成，且拓扑结构及属性随时间不断演化，其所蕴含的异质性和动态性可以通过动态异质图数据完整描述。为缓解现有动态异质图表示学习模型存在时空信息融合较为粗糙，及其监督学习范式强依赖于人工标签的问题，该文提出了掩码自编码器增强的动态异质图表示学习模型：首先通过多层次注意力结构融合异质空间信息并进行跨快照的时间信息聚合，其次利用掩码自编码器的重建损失丰富节点的表示信息。该文模型在多个真实世界数据集的链路预测任务中相较基线模型均有1%-4%的AUC(Area Under the Receiver Operating Characteristic Curve)提升。该文模型为动态异质图表示学习提供了一个有效的自监督框架，有助于更精确地捕捉真实网络中的异质信息与动态演化规律。

关键词: 表征学习, 动态异质图, 掩码自编码器, 自监督学习, 链路预测