Pre-Conference Talk by LIU Ran | Diversified and Adaptive Negative Sampling for Knowledge Graph

In knowledge graph embedding, both positive and negative triplets impact model performance. Since knowledge graphs are sparse and often lack explicit negative labels, negative samples are typically generated by randomly replacing entities in positive triplets. Ideally, these negatives should be informative, but existing methods often overlook diversity and adaptiveness, limiting their effectiveness. We propose DANS (Diversified and Adaptive Negative Sampling), a generative adversarial approach that improves sampling through two key innovations; 1. A two-way generator for producing more diverse negative triplets and 2. An adaptive mechanism that localizes sampling for specific entities and relations, yielding fine-grained examples. Experiments on three benchmark datasets demonstrate the effectiveness of DANS through both quantitative and qualitative results.

This is a Pre-Conference talk for The 9th International Workshop on Graph Data Management and Analysis (GDMA 2025).

LIU Ran is a PhD Candidate in Computer Science at the SMU School of Computing and Information Systems, supervised by Assistant Prof. FANG Yuan.  His research is focused on Artificial Intelligence & Data Science and the main research interest lies in the topic of knowledge graph representation learning.