Next-destination prediction in continuous-time dynamic graphs (CTDGs) suffers from evaluation bias where non-uniform negative distributions and finite candidate sets alter Bayes-optimal rankings. Testing across LastFM, MOOC, Reddit, and Wikipedia shows that sampling methods directly influence measured model performance and module effects.
HOW THIS AFFECTS YOU
●
researcherYou should account for sampler-induced bias to ensure your CTDG model comparisons are valid.