Community-oriented multi-scale heterogeneous community detection using weighted positives and debiased negatives

• Published in: Knowledge-based systems 2025-02, Vol.310, p.112934, Article 112934
• Main Authors: Liu, Fangfang, Li, Chunjie, Wu, Bo
• Format: Article
• Language: English
• Subjects: Community detection; Community semantic; Heterogeneous network; Meta-path interaction; Positives and negatives sampling
• ISSN: 0950-7051
• DOI: 10.1016/j.knosys.2024.112934

Community detection aims to identify groups of closely connected nodes in complex networks. While community detection methods for homogeneous networks are well-developed, they struggle with higher-order semantics and scalability in heterogeneous networks. Moreover, existing heterogeneous community detection approaches typically focus on node representations followed by clustering, often neglecting important community structure semantics and higher-order mate-path interaction semantics. Contrastive learning is an effective measure for community detection but is prone to introducing noise such as False Positives (FPs) and False Negatives (FNs), with prior work failing to address both issues simultaneously. To tackle these challenges, we propose a Community-oriented Multi-scale heterogeneous Community detection using Weighted positives and Debiased negatives (CMCWD) approach. This method constructs Node-Wise Heterogeneous information Modeling (NWHM) and Community-Wise Heterogeneous information Modeling (CWHM) using contrastive learning and supplemented with generative learning to emphasize structural information between nodes and avoid neglecting important community structures. Specifically, NWHM captures multi-scale local and higher-order meta-path interaction information from both a local topological view and a meta-path distillation view, considering both the independence and interactivity of meta-paths. Meanwhile, CWHM captures community structure semantics using Non-negative matrix factorization-based generative learning. Additionally, CMCWD introduces two denoising strategies – Robust Cross-view Weighted contrastive Mechanism (RCWM) and Community-Oriented Debiased contrastive Mechanism (CODM) – to simultaneously mitigate FPs and FNs in community detection. We conduct extensive experiments on six public datasets to validate the effectiveness of CMCWD. The results demonstrate that CMCWD outperforms other mainstream methods, achieving a 9.98% higher NMI compared to the second-best unsupervised method on the AMINER, underscoring its effectiveness.