| |
 Graph Perturbations for Robust Knowledge Discovery and Retrieval |  | XIAO Hanhua PhD Candidate
School of Computing and Information Systems
Singapore Management University
| Research Area Dissertation Committee Research Advisor Committee Members External Member - Roy Ka-Wei LEE, Cheng Tsang Man Early Career Chair Professor, Information Systems Technology and Design Pillar, Singapore University of Technology and Design
|
| | Date 21 November 2025 (Friday) | Time 3:30pm - 4:30pm | Venue Meeting room 5.1,
Level 5
School of Computing and Information Systems 1,
Singapore Management University,
80 Stamford Road
Singapore 178902 | Please register by 19 November 2025. We look forward to seeing you at this research seminar. 
|
|
|
| | ABOUT THE TALK Graph perturbation studies how small structural edits, e.g., adding or deleting edges, affect graph properties and downstream tasks. While prior work largely targets global statistics or model outputs, robustness for knowledge discovery and information retrieval on attributed graphs is less explored. This dissertation introduces formulations and algorithms that generate and leverage perturbations to make these tasks more reliable.
First, we assess the robustness of outstanding facts on knowledge graphs. We formalize two perturbation types: (i) context‑entity swaps and (ii) plausible data insertions and estimate robustness via a perturbation‑relevance distribution, with powerful relevance‑based pruning and importance‑sampling estimators.
Second, we study subgraph counting under adversarial edge additions. We prove the kSub problem is NP‑hard and inapproximable, derive the topkSub relaxation, and design connected/disconnected processing with pruning and unbiased samplers, enabling practical stress tests that expose weaknesses in GNN‑based counters.
Finally, for robust information retrieval via retrieval‑augmented generation (RAG), we present GBRAG, a training‑free graph‑bandit framework that treats chunks as bandit arms and updates a query‑conditioned priority map using two LLM‑scored rewards: immediate relevance and marginal information gain, yielding consistent gains over strong graph‑RAG baselines.
Across extensive evaluations, the methods are efficient, theoretically grounded, and effective for robust knowledge discovery and information retrieval. | | | SPEAKER BIOGRAPHY XIAO Hanhua is a Ph.D. candidate in Computer Science at Singapore Management University, under the supervision of Associate Professor Li Yuchen. His research focuses on graph mining, with particular interests in knowledge graph mining, graph perturbation, and large language model (LLM)-based graph analytics. Before joining SMU, Hanhua received his B.Sc. and M.Sc. degrees in Electronic Engineering from the Beijing University of Posts and Telecommunications and the University of Southern California in 2017 and 2019, respectively. In his leisure time, he enjoys cycling and cooking. |
|
