PhD Dissertation Defense by LIU Ran | Addressing Sparsity For Knowledge Graph Completion: Data And Model Perspectives

Addressing Sparsity For Knowledge Graph Completion:
Data And Model Perspectives

ABOUT THE TALK

This dissertation addresses data sparsity in KGC through complementary data- and model-level solutions. It introduces (1) Diversified and Adaptive Negative Sampling (DANS) to produce more informative negatives for supervised learning, (2) FusionAdapter for modality-preserving, parameter-efficient multimodal fusion to mitigate long-tail relation issues, and (3) RelAdapter, a context-aware adapter that enables relation-specific adaptation under distribution shift. In addition, an ongoing line of work is outlined as future research, which explores bridging knowledge graphs with large language models through interpretable tokenization. Overall, these contributions provide a cohesive path toward more robust and generalizable knowledge graph completion.

SPEAKER BIOGRAPHY

LIU Ran is a PhD candidate in Computer Science at Singapore Management University, supported by the ASTAR Graduate Scholarship. His research focuses on knowledge graphs, few-shot learning, and multimodal fusion, and he has published in venues such as EMNLP and GDMA. Previously, he worked as a research attachment at ASTAR I2R and completed a data science internship at Point72, where he applied graph-based machine learning to financial problems. He also has in the Inland Revenue Authority of Singapore, where he worked on tax analytics, optimization, and automation projects. In his free time, he enjoys watching movies, fishing, and traveling.