Graph Prompt Clustering.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Unlabeled graph-structured data is abundant, driving interest in graph-level clustering.
• Existing methods often overlook distinct data distributions across datasets.
• Adapting models to diverse graph datasets without prior knowledge remains a challenge.

Purpose of the Study:

• To propose a novel Graph Prompt Clustering (GPC) method.
• To address the challenge of clustering multiple graph-level datasets with varying distributions.
• To develop a generalizable model for graph clustering.

Main Methods:

• A two-module approach: graph model pretraining and prompt-based finetuning.
• Pretraining utilizes mutual information maximization and self-supervised clustering regularization.
• Finetuning employs frozen pretrained parameters with learnable prompt vectors for adaptation.

Main Results:

• GPC demonstrates impressive generalization capability across six benchmark datasets.
• The method effectively adapts to different target graph-level datasets.
• Experimental results show GPC outperforms state-of-the-art methods.

Conclusions:

• GPC offers an effective and generalizable solution for graph-level clustering.
• The prompt-based adaptation mechanism successfully handles diverse data distributions.
• This approach advances the field of unsupervised learning on graph data.