Scholarlink

We propose CI-FSA, a scalable algorithm for extracting Collective Influence-Focal Sets (CI-F Sets)—groups of nodes that are both structurally cohesive and propagation-aware. Bridging community detection and influence maximization, CI-FSA identifies influential subnetworks in complex systems. Evaluations on four open-source social networks demonstrate that CI-FSA outperforms baselines across various metrics, including giant component ratio, ASPL increase, connectivity loss, and network efficiency drop. Additionally, complexity analysis reveals its scalability compared to existing methods. We demonstrated the applicability of our method through an experiment using real-world datasets, comprising a multi-platform graph from the 2024 Taiwan Election and Enron email communication network. Our findings highlight that influence often emerges from tightly connected groups rather than top-ranked individuals. The approach yields interpretable clusters functional in applications such as manipulation detection and narrative mapping. Future work will explore adaptive initialization, parameter ablation, and extensions to temporal and multiplex networks.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

CI-FSA: Toward Scalable Discovery of Influential Groups in Social Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ridwan Amure,
  • Nitin Agarwal

摘要

We propose CI-FSA, a scalable algorithm for extracting Collective Influence-Focal Sets (CI-F Sets)—groups of nodes that are both structurally cohesive and propagation-aware. Bridging community detection and influence maximization, CI-FSA identifies influential subnetworks in complex systems. Evaluations on four open-source social networks demonstrate that CI-FSA outperforms baselines across various metrics, including giant component ratio, ASPL increase, connectivity loss, and network efficiency drop. Additionally, complexity analysis reveals its scalability compared to existing methods. We demonstrated the applicability of our method through an experiment using real-world datasets, comprising a multi-platform graph from the 2024 Taiwan Election and Enron email communication network. Our findings highlight that influence often emerges from tightly connected groups rather than top-ranked individuals. The approach yields interpretable clusters functional in applications such as manipulation detection and narrative mapping. Future work will explore adaptive initialization, parameter ablation, and extensions to temporal and multiplex networks.