Information bottleneck-driven prompt on graphs for unifying downstream few-shot classification tasks

Inspired by the success of prompt in natural language processing, the graph prompt-based methods are proposed to solve the classification tasks under the conditions with limited instances. Recent graph prompt-based methods typically employ link prediction as the pre-training task, which brings a gap between the pre-training task and the downstream classification task, thus introducing irrelevant and noisy features to the downstream tasks. To tackle this issue, we propose a framework called Diffused Graph Prompt (Di-Graph), which consists of three major components, i.e., the diffused pre-training, a graph prompt layer, and an information bottleneck optimizer. Specifically, the diffused pre-training aims to obtain the stable node features with a diffusion process, mitigating the gap between the pre-training tasks and the downstream tasks. The graph prompt layer enhances the pre-trained model to leverage its knowledge via capturing both the structural and node features in graphs. The information bottleneck optimizer helps the model discard redundant features by retaining the minimal sufficient statistic of the input data. Extensive experimental results on five public graph datasets demonstrate that our Di-Graph model surpasses the state-of-the-art model in terms of accuracy for both node and graph classification tasks. In particular, for graph-level tasks, Di-Graph achieves an average performance gain of 6.50% over the previous best model (GraphCL) on BZR dataset. For node-level tasks, Di-Graph achieves a 2.05% improvement over the best baseline (GraphCL) on PROTEINS dataset.