In recent years, prompt tuning has set off a research boom in the adaptation of pre-trained models. In this paper, we propose Graph Prompt as an efficient and effective alternative to full fine-tuning for adapting the pre-trianed GNN models to downstream tasks. To the best of our knowledge, we are the first to explore the effectiveness of prompt tuning on existing pre-trained GNN models. Specifically, without tuning the parameters of the pre-trained GNN model, we train a task-specific graph prompt that provides graph-level transformations on the downstream graphs during the adaptation stage. Then, we introduce a concrete implementation of the graph prompt, called GP-Feature (GPF), which adds learnable perturbations to the feature space of the downstream graph. GPF has a strong expressive ability that it can modify both the node features and the graph structure implicitly. Accordingly, we demonstrate that GPF can achieve the approximately equivalent effect of any graph-level transformations under most existing pre-trained GNN models. We validate the effectiveness of GPF on numerous pre-trained GNN models, and the experimental results show that with a small amount (about 0.1% of that for fine-tuning ) of tunable parameters, GPF can achieve comparable performances as fine-tuning, and even obtain significant performance gains in some cases.
View on arXiv