Prof Nitesh Chawla, University of Notre Dame
Prof. Chanyoung Park, KAIST
Detecting Change, Not Just State: Change-Aware Graph Learning for Evolving Fraud and Abuse
Fraud and abuse are inherently dynamic: benign entities can become risky, yet most detection models focus on static states rather than transitions. In this talk, I introduce label shifters—entities whose labels change over time—and show that dynamic graph neural networks systematically struggle with them due to embedding inertia, which keeps representations anchored to stale neighborhood signals. I will then present CHASE, a model-agnostic framework that explicitly detects change and enables models to override outdated information. Our results show consistent improvements on shifting entities while preserving performance on stable ones. The broader message is simple: for fraud and abuse, we should ask not only “Who is risky?” but “Who just became risky?”
Prof. Jaemin Yoo, Seoul National University
Saurabh Nagrecha & Shalini Ghosh, Google