Reproduce “PromptDyG: Test-Time Prompt Adaptation on Dynamic Graphs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Guoguo Ai, Chaoxi Niu, Hui Yan, Joey Tianyi Zhou, Yew-Soon Ong, Guansong Pang
PromptDyG introduces a label-free test-time adaptation method for temporal link prediction on discrete-time dynamic graphs. It freezes the pretrained dynamic GNN and updates only a lightweight input prompt using feature-entropy minimization, achieving the best MRR on six benchmarks while improving the average MRR from 0.2627 to 0.2794 over the Roland-GRU backbone with lower adaptation overhead than competing approaches.
Key insights
- Structural shifts at test time can be handled by modifying the model input rather than its temporal parameters, preserving information accumulated from earlier graph snapshots.
- Entropy minimization is used to align test-time node representations with dominant structural patterns, increasing similarity among likely linked nodes and separating unlikely pairs.
- The approach treats the final projection directions as evolutionary prototypes; under assumptions about prototype separation and temporal-spatial consistency, sharpening prototype assignments should widen the link-prediction margin.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.