丁香婷婷激情综合俺也去

发表期刊：Neuro computing

分区：厂颁滨1区

发表时间：2025年8月19日

第一作者：李有红

中文摘要：随着组织网络日益动态化和多模态化，实时社区检测已成为跨国重组与应急响应的关键技术。现有方法在事件级联建模、跨模态语义对齐以及隐私保护与检测精度平衡方面面临挑战。本文提出联邦动态跨模态社区检测框架（FedDyCM），其核心创新主要体现在：(1)将霍克斯过程与时序图神经网络相结合，通过自激强度矩阵量化事件级联以建模突发性结构变化；(2)采用门控跨模态注意力机制，通过上下文感知注意力动态调整文本、拓扑和行为特征权重，在危机场景中自适应提升关键模态权重以增强语义对齐；(3)构建联邦隐私对齐机制，将局部差分隐私与Procrustes嵌入对齐相结合，在确保全局语义一致性的同时，将隐私攻击率控制在<5 %；(4)通过熵约束公平性机制限制最大社区规模，防止资源垄断。实验结果表明：该框架满足级联效应、跨模态融合、分布均衡和隐私保护等标准，在隐私约束下实现了实时社区划分。FedDyCM模型动态模块化得分为0.50 &plusmn; 0.01，将社区重组检测的F1值提升至0.88 &plusmn; 0.01，并使归一化互信息（NMI）较基线模型提高13.3个百分点。该框架为跨国组织管理提供高效、安全且公平的决策支持，推动动态图神经网络在复杂现实场景中的实际应用。

英文摘要：As organizational networks grow increasingly dynamic and multimodal, real-time community detection has become critical for multinational restructuring and emergency response. Existing methods face challenges in modeling event cascades, resolving cross-modal semantic alignment, and balancing privacy protection with detection accuracy. This paper presents Federated Dynamic Cross-Modal Community Detection Framework(FedDyCM),the core innovation is mainly reflected in: (1) integration of Hawkes processes with temporal graph neural networks to quantify event cascades via self-exciting intensity matrices for modeling sudden structural shifts; (2)gated cross-modal attention that dynamically weights text, topological, and behavioral features through contextaware attention to enhance semantic alignment by adaptively elevating critical modality weights during crisis scenarios; (3) a federated privacy-alignment mechanism combining local differential privacy with Procrustes embedding alignment to preserve data sovereignty while ensuring global semantic consistency and limiting privacy attack rates to <5 %; and (4) entropy-constrained fairness to prevent resource monopolization by limiting maximum community size. Experimental results show: This framework satisfies the criteria of cascade effects,cross-modal fusion, balanced distribution, and privacy preservation, achieving real-time community partitioning under privacy constraints. FedDyCM achieves a dynamic modularity score of 0.50 &plusmn; 0.01, elevates the F1-score for community reorganization detection to 0.88 &plusmn; 0.01, and improves Normalized Mutual Information (NMI) by 13.3 percentage points compared to baseline models. This framework provides efficient, secure, and equitable decision support for multinational organizational management, advancing practical application of dynamic graph neural networks in complex real-world scenario.