Prompt Tuning for Graph Neural Networks

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.

DropMessage: Unifying Random Dropping for Graph Neural Networks

Graph Neural Networks (GNNs) are powerful tools for graph representation learning. Despite their rapid development, GNNs also faces some challenges, such as over-fitting, over-smoothing, and non-robustness. Previous works indicate that these problems can be alleviated by random dropping methods, which integrate noises into models by randomly masking parts of the input. However, some open-ended problems of random dropping on GNNs remain to solve. First, it is challenging to find a universal method that are suitable for all cases considering the divergence of different datasets and models. Second, random noises introduced to GNNs cause the incomplete coverage of parameters and unstable training process. In this paper, we propose a novel random dropping method called DropMessage, which performs dropping operations directly on the message matrix and can be applied to any message-passing GNNs. Furthermore, we elaborate the superiority of DropMessage: it stabilizes the training process by reducing sample variance; it keeps information diversity from the perspective of information theory, which makes it a theoretical upper bound of other methods. Also, we unify existing random dropping methods into our framework and analyze their effects on GNNs. To evaluate our proposed method, we conduct experiments that aims for multiple tasks on five public datasets and two industrial datasets with various backbone models. The experimental results show that DropMessage has both advantages of effectiveness and generalization.