Graph neural networks (GNNs) have made remarkable advancements in processing graph-structured data with all nodes and edges belonging to the same type. However, various types of node and relations exist in heterogeneous information networks (HINs), and due to this, HINs contain rich structural and semantic information. To tackle this heterogeneity, existing methods usually apply several well-designed metapaths to HINs to obtain the corresponding homogeneous subgraphs. However, these methods either fail to capture the interconnections between the same nodes in different subgraphs or require qualified labels. To address these issues, we propose a new multi-aspect self-supervised learning (SSL) framework for HIN representation in an unsupervised manner: (1) we design a new contrastive learning model to capture the similarities between the same nodes in different homogeneous subgraphs, and (2) we maximize the mutual information between the local patches and the global representation in one subgraph. Extensive experiments on various downstream tasks demonstrate the superiority of our model in comparison to the existing state-of-the-art methods.

图神经网络 (GNN) 在处理所有节点和边属于同一类型的图结构数据方面取得了显着进步。然而，异构信息网络（HIN）中存在各种类型的节点和关系，因此，HIN包含丰富的结构和语义信息。为了解决这种异质性，现有方法通常将几个设计良好的元路径应用于 HIN 以获得相应的同质子图。然而，这些方法要么无法捕获不同子图中相同节点之间的互连，要么需要合格的标签。为了解决这些问题，我们提出了一种新的多方面自监督学习 (SSL) 框架，用于以无监督方式进行 HIN 表示：(1) 我们设计了一个新的对比学习模型来捕捉不同同构子图中相同节点之间的相似性，以及 (2) 我们最大化局部块和一个子图中的全局表示之间的互信息。对各种下游任务的大量实验证明了我们的模型与现有最先进方法相比的优越性。