关键词: 图挖掘, 图提示学习, 图神经网络, 少样本, 聚类

Abstract:  In graph mining tasks, prototype-based graph prompt learning has been widely regarded as an effective means to improve the performance of graph data analysis. However, in the scenario of fest-sample node classification, existing methods have problems such as inaccurate construction of class prototypes due to insufficient utilization of unlabeled data and inadequate utilization of graph topological structure information. These deficiencies limit the effectiveness of graph prompt learning methods in downstream tasks. To this end, a graph prompt learning method PNCD-GP that integrates the cluster distribution of all nodes is proposed, aiming to improve the performance and accuracy of the analysis by fully utilizing the cluster distribution and topological structure information of unlabeled data. In the pre-training stage, predictive masks and preserving graph node clustering are adopted as optimization strategies to learn discriminative representations and narrow the gap between upstream and downstream tasks. During the prompt fine-tuning model learning stage, class prototype virtual nodes are introduced into the original graph as prompts, and high-order information is utilized to enhance the topological structure, thereby improving the model's understanding and utilization of the graph structure. Learn cues by maintaining the cluster distribution of unlabeled samples and labeled nodes. This method can construct more accurate prototype vectors and classify nodes by utilizing the similarity between class prototypes and node representations. Experimental results on multiple public graph datasets show that the PNCD-GP method has significant advantages in both efficiency and accuracy.

Key words: graph mining, graph prompt learning, graph neural network, few-shot, clustering