Approximate personalized propagation for unsupervised embedding in heterogeneous graphs

• Main Authors: Chen, Yibi, Hu, Yikun, Li, Keqin, Yeo, Chai Kiat, Li, Kenli
• Other Authors: School of Computer Science and Engineering
• Format: Article
• Language: English
• Published: 2022
• Subjects: Engineering::Computer science and engineering; Approximate Personalized Propagation; Heterogeneous Graphs
• Online Access: https://hdl.handle.net/10356/163878
• Institution: Nanyang Technological University

Graphs are effective for representing various relationships in the real world and have been successfully applied in many areas, such as publication citations and movie networks. Compared to homogeneous graphs (i.e., nodes and edges of a single relation type), heterogeneous graphs have heterogeneity and richer information (i.e., nodes and edges of different relation types). How to tackle complex non-pairwise graph-structured data and model various relation-types is a daunting challenge for heterogenous graphs. However, the existing unsupervised methods focus on node attribute learning, while node neighborhood information utilizes very limited because they only consider node propagation that is within few steps. In this paper, we propose an unsupervised method, called APPTE, that models adequate node neighborhood information in local context, and captures the global neighborhood information. Meanwhile, our method considers the robustness and generalization ability. Specifically, we construct approximate personalized propagation in local context to utilize an infinite number of neighborhood aggregation layers for extending node neighborhood propagation range, and then fuse these local context to capture global neighborhood information. Additionally, we improve the robustness and generalization ability of model, employing throwedge to increase the randomness and diversity of the graph connections by randomly deleting a part of edges. The experimental results on three benchmark datasets containing heterogeneous graphs demonstrate that our proposed method is superior to the available state-of-the-art methods.