| |
 | | From Suspicious Nodes to Multi-Agent Failures: Learning to Detect, Generalize, and Diagnose Unusual Behaviors |  | QIAO Hezhe PhD Candidate
School of Computing and Information Systems
Singapore Management University
FULL PROFILE |
Research Area - Artificial Intelligence & Data Science
- Machine Learning & Intelligence
Dissertation Committee | Advisor: | | | Members: | | | | | | External Members: | LI Xiaoli, Professor and Head, Information Systems Technology and Design (ISTD) Pillar, Singapore University of Technology and Design |
| | Date 16 November 2026 (Monday) Time 3:00pm - 4:00pm Venue Meeting room 4.4, Level 4
School of Computing and Information Systems 1,
Singapore Management University,
80 Stamford Road,
Singapore 178902 Please register by 15 November 2026. We look forward to seeing you at this research seminar. 
|
| ABOUT THE TALK Detecting unusual behavior is essential for safeguarding interconnected systems, from social and financial networks to large language model (LLM)–based multi-agent systems. This thesis addresses limited anomaly supervision, generalization across graph domains, and failure diagnosis through four complementary studies.
First, TAM enables unsupervised graph anomaly detection by modeling one-class homophily and exploiting the structural and contextual consistency of normal nodes without requiring labels. Second, GGAD addresses semi-supervised detection by generating informative pseudo-anomalies and combining anomaly-aware representation learning with a one-class objective, effectively using limited normal labels.
Third, AnomalyGFM supports zero-shot and few-shot graph anomaly detection through residual representations, normality and abnormality prototypes, and localized subgraph inference. It transfers patterns learned from a source graph to unseen target graphs with no or limited target labels.
Finally, VerifyMAS extends the investigation to failure diagnosis in LLM-based multi-agent systems. Through error-first reasoning and evidence-based hypothesis verification, it analyzes complete interaction trajectories to identify predefined failure types and associate them with responsible agents.
Together, these studies advance anomaly detection under limited supervision, transfer across graph domains, and retrospective failure attribution in collaborative AI systems. Their common principle is to assess unusual behavior within its structural or interaction context. | ABOUT THE SPEAKER Hezhe QIAO is a final-year PhD candidate in Computer Science at Singapore Management University, supervised by Professor Guansong Pang. His research focuses on trustworthy and generalizable AI, particularly graph anomaly detection, foundation models, and the reliability of LLM-based multi-agent systems. His related work has been published at leading venues, including NeurIPS x4, ICML x2, KDD x2, IJCAI, and IEEE TKDE. He has also presented tutorials at IJCAI2025 and AAAI2026, sharing research advances with the broader AI community. In recognition of his research achievements, he has received the SMU Presidential Doctoral Fellowship and been named to the SCIS Dean’s List. |
|
