Subgraph-level Universal Prompt Tuning

In the evolving landscape of machine learning, the adaptation of pre-trainedmodels through prompt tuning has become increasingly prominent. This trend isparticularly observable in the graph domain, where diverse pre-trainingstrategies present unique challenges in developing effective prompt-basedtuning methods for graph neural networks. Previous approaches have beenlimited, focusing on specialized prompting functions tailored to models withedge prediction pre-training tasks. These methods, however, suffer from a lackof generalizability across different pre-training strategies. Recently, asimple prompt tuning method has been designed for any pre-training strategy,functioning within the input graph's feature space. This allows it totheoretically emulate any type of prompting function, thereby significantlyincreasing its versatility for a range of downstream applications.Nevertheless, the capacity of such simple prompts to fully grasp the complexcontexts found in graphs remains an open question, necessitating furtherinvestigation. Addressing this challenge, our work introduces theSubgraph-level Universal Prompt Tuning (SUPT) approach, focusing on thedetailed context within subgraphs. In SUPT, prompt features are assigned at thesubgraph-level, preserving the method's universal capability. This requiresextremely fewer tuning parameters than fine-tuning-based methods, outperformingthem in 42 out of 45 full-shot scenario experiments with an average improvementof over 2.5achieving an average performance increase of more than 6.6