Subgraph-level universal prompt tuning

초록

In the evolving landscape of machine learning, adapting pre-trained models through prompt tuning has become increasingly significant, particularly in the graph domain. Diverse pre-training strategies for graph neural networks pose unique challenges for effective prompt-based tuning. Prior approaches have been limited, focusing on specialized prompting functions tailored to edge prediction tasks, which lack generalizability. A recent simple prompt tuning method works across various pre-training strategies by functioning within the input graph's feature space. Theoretically, it can emulate any prompting function, enhancing versatility for downstream applications. However, its ability to grasp complex graph contexts remains uncertain, necessitating further investigation. Addressing this, we introduce Subgraph-level Universal Prompt Tuning (SUPT), focusing on detailed subgraph contexts while preserving universal applicability. SUPT requires significantly fewer tuning parameters than fine-tuning methods, outperforming them in 42 out of 45 full-shot scenario experiments with an average improvement of over 2.5%. In few-shot scenarios, SUPT excels in 41 out of 45 experiments, achieving an average performance increase of more than 6.6%.

키워드