Reproduce “MultiGPrompt for Multi-Task Pre-Training and Prompting on 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
Xingtong Yu, Chang Zhou, Yuan Fang, Xinming Zhang
A framework called MultiGPrompt leverages multi-task pre-training and a dual-prompt mechanism for graph neural networks to enhance knowledge acquisition and transfer in low-data settings. It consistently outperformed state-of-the-art baselines on six datasets for few-shot node and graph classification tasks, demonstrating improved accuracy and parameter efficiency.
Key insights
- Task-specific "pretext tokens" can effectively facilitate synergistic optimization of diverse pretext tasks during multi-task pre-training by layer-wise modification of graph encoder outputs.
- A dual-prompt mechanism, combining composed prompts and open prompts, enables flexible and comprehensive transfer of both task-specific and global inter-task knowledge to downstream tasks.
- Leveraging multiple pretext tasks during pre-training yields a more comprehensive and robust set of pre-trained graph representations compared to relying on a single pretext task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.