报告题目：Uncertainty-Aware Deep Graph Clustering and its application in Spatial Transcriptomics Data

报 告 人：易思宇助理研究员（四川大学）

报告时间：2025年11月8日（星期六）9:50-10:30

报告地点：数学大楼报告厅三（814）

参加人员：教师、研究生、本科生

报告摘要：Deep graph clustering (DGC) has gained significant attention due to its potential in partitioning nodes in graphs into meaningful clusters. However, existing DGC methods often overlook the impact of inherent data noise, which introduces aleatoric uncertainty and degrades clustering performance. In this paper, we propose a novel uncertainty-aware graph clustering framework, UGC, to address this challenge. UGC refines original graph structure by enhancing informative connections and suppresses spurious edges, and models distributional representations to capture uncertainty caused by both feature- and structure-level ambiguity. This uncertainty is then incorporated into the training process through contrastive learning, guided by high-confidence topology and clustering information. Moreover, UGC also introduces consistency learning to facilitate the alignment of distributional semantics across dual views, promoting view-specific noise suppression. Finally, an uncertainty-aware representation fusion mechanism is designed to further mitigate the impact of noise or contradictory information for final clustering. Extensive experiments on benchmark datasets demonstrate that UGC consistently outperforms state-of-the-art methods, achieving superior graph clustering performance. Furthermore, we extend UGC to domain identification in spatial transcriptomic data, where experimental validation and analysis further confirm its effectiveness and robustness.

报告人简介：易思宇，四川大学数学学院助理研究员，入选国家博士后创新人才支持计划（国资计划A档）。于2024年在南开大学统计与数据科学学院获博士学位，研究方向为图机器学习、大数据子抽样、AI4Science等。主持国家自然科学基金青年科学基金项目、中国博士后科学基金面上项目、四川省自然科学基金青年科学基金项目。研究成果发表于J. Comput. Graph. Stat., Stat. Comput., IEEE TNNLS, IEEE TMM, IEEE TBD, ICML, NeurIPS, AAAI, IJCAI等统计学和机器学习领域的高水平期刊和会议上。