数学院综合集成与知识科学研究组 Home    |    Contact   |    中文   |   ISS   |    CAS
Some Results on Signed Graph Representati...
作者:Jiamou Liu(The University of Auckland) 来源 : 中科院数学院南楼N205 时间:2025-07-16 字体<    >
   
题目:Some Results on Signed Graph Representation Learning
时间:2025年7月16日(周三)9:30-10:30
地点:中科院数学院南楼N205
报告人:Jiamou Liu(The University of Auckland)

报告摘要:
The unique challenges of signed graphs, where both positive and negative relationships coexist, demand principled solutions that go beyond conventional graph representation learning methods. In this talk, I present a unified research trajectory that begins with a general-purpose robustness framework and culminates in a task-specific contrastive learning approach for structured signed graphs. I first introduce RSGNN, a model-agnostic framework designed to denoise signed graphs by leveraging theoretical insights from an extended Weisfeiler-Lehman test and structural balance theory. RSGNN effectively mitigates the effects of edge noise, enabling existing SGNN models to recover robust representations. Building on this foundation, I present SBGCL, a novel contrastive learning paradigm tailored for signed bipartite graphs. Through a two-level augmentation strategy and a multi-perspective contrastive objective, SBGCL captures both explicit interactions and implicit intra-set relations, improving resilience and interpretability.  Together, these works offer a comprehensive approach to robust and expressive signed graph representation learning, relevant to researchers and practitioners working in trust analysis, recommender systems, and social network modeling. 

符号图(包含正负关系)带来的独特挑战,亟需超越传统图表示学习方法的、基于原理的解决方案。在本次报告中,我将介绍一个统一的研究路径:它始于一个通用的鲁棒性框架,并最终发展为针对结构化符号图的、面向特定任务的对比学习方法。我首先介绍RSGNN,这是一个模型无关的框架,旨在通过利用扩展的魏菲勒-莱曼测试(Weisfeiler-Lehman test)和结构平衡理论的理论洞见来对符号图进行去噪。RSGNN 有效减轻了边噪声的影响,使现有的符号图神经网络(SGNN)模型能够恢复鲁棒表示。在此基础上,我提出SBGCL,这是一种专为符号二部图设计的新型对比学习范式。通过双层增强策略和多视角对比目标,SBGCL 能够同时捕捉显式的交互关系和隐式的集合内关系,从而提升了模型的鲁棒性和可解释性。总而言之,这些工作共同为鲁棒且富有表达力的符号图表示学习提供了一套全面的方法,对从事信任分析、推荐系统和社会网络建模等领域的研究与实践人员具有参考价值。
 
报告人简介: 
Jiamou Liu is an Associate Professor at the School of Computer Science, The University of Auckland, New Zealand. His research spans across several directions in artificial intelligence, in particular, multi-agent systems, algorithmic mechanism design, data marketplaces, and natural language processing. He has recently made contributions to fairness-aware auction mechanisms and privacy-preserving data trading protocols. Dr. Liu has authored over 140 peer-reviewed publications in top-tier venues such as AAAI, AAMAS, IJCAI, WWW, SIGIR, and ICML. He is a recipient of multiple research grants, including Marsden Fund projects and national funding from New Zealand.
相关附件
相关文档

CAS,Research Group of Meta-Synthesis and Knowledge Science
京ICP备05002806号-6  文保网安备案号 1101080081 邮箱: mcs@iss.ac.cn
电话:+86 10 82541801