异质网络中融合多种语义关系的高效社区搜索方法

Efficient community search with multiple semantic relationships in heterogeneous information networks

关键词:  异质信息网络  社区搜索  多种语义关系  离线学习  在线搜索

Key words:  heterogeneous information networks  community search  multiple semantic relationships  offline learning  online query

摘要:

为解决异质信息网络中现有社区搜索方法存在的局限性,本文提出了一种融合多种语义关系的异质信息网络社区搜索方法,采用高效的“离线学习在线搜索”策略,其核心在于:利用语义注意力机制自适应学习不同元路径对目标社区凝聚性的权重贡献,精准量化语义差异；再结合网络结构与节点属性特征度量节点相关性,定位社区成员。离线阶段预训练节点社区关联模型,生成节点归属各类社区的概率分布向量；在线阶段基于预计算结果快速响应社区搜索。此策略既可保持学习模型的灵活性,有效捕捉异质网络语义与属性,又将主要计算负担置于离线阶段,显著提升查询效率,尤其适用于高频场景。在多个真实数据集上的验证实验表明,本方法在社区有效性（语义相关性、结构凝聚性、属性一致性）和查询效率上均显著优于现有主流方法。

Abstract:

To address key limitations in existing community search methods for heterogeneous information networks (HINs) , this paper proposes a community search method for HINs that integrates multiple semantic relationships, employing an efficient “offline learning, online search” strategy. Its core lies in: adaptively learning the weight contributions of different meta-paths to the target community cohesiveness using a semantic attention mechanism to precisely quantify semantic differences; and subsequently measuring node relevance by combining network structure and node attribute features to locate community members. In the offline phase, a node-community association model is pre-trained to generate probability distribution vectors indicating node affiliation across various communities. In the online phase, community search is rapidly responded to based on precomputed results. This strategy maintains the flexibility of the learning model to effectively capture heterogeneous network semantics and attributes, while shifting the main computational burden to the offline phase, significantly improving query efficiency, making it particularly suitable for high-query-frequency scenarios. Experiments on multiple real-world HIN datasets demonstrate that our method significantly outperforms existing mainstream methods in both community effectiveness (semantic relevance, structural cohesiveness, attribute consistency) and query efficiency.