FCS-HGNN: Flexible Multi-type Community Search in Heterogeneous Information Networks
Guoxin ChenFangda GuoYongqing WangYanghao Liu and Peiying YuHuawei ShenXueqi Cheng
Guoxin ChenFangda GuoYongqing WangYanghao Liu and Peiying YuHuawei ShenXueqi Cheng
Nov 2023
0被引用
1笔记
摘要原文
Community Search (CS), a crucial task in network science, has attracted considerable interest owing to its prowess in unveiling personalized communities, thereby finding applications across diverse domains. Existing research primarily focuses on traditional homogeneous networks, which cannot be directly applied to heterogeneous information networks (HINs). However, existing research also has some limitations. For instance, either they solely focus on single-type or multi-type community search, which severely lacking flexibility, or they require users to specify meta-paths or predefined community structures, which poses significant challenges for users who are unfamiliar with community search and HINs. In this paper, we propose an innovative method, FCS-HGNN, that can flexibly identify either single-type or multi-type communities in HINs based on user preferences. We propose the heterogeneous information transformer to handle node heterogeneity, and the edge-semantic attention mechanism to address edge heterogeneity. This not only considers the varying contributions of edges when identifying different communities, but also expertly circumvents the challenges presented by meta-paths, thereby elegantly unifying the single-type and multi-type community search problems. Moreover, to enhance the applicability on large-scale graphs, we propose the neighbor sampling and depth-based heuristic search strategies, resulting in LS-FCS-HGNN. This algorithm significantly improves training and query efficiency while maintaining outstanding community effectiveness. We conducted extensive experiments on five real-world large-scale HINs, and the results demonstrated that the effectiveness and efficiency of our proposed method, which significantly outperforms state-of-the-art methods.
