Reliable and Compact Graph Fine-Tuning Via Graph Sparse Prompting.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Science

Background:

• Graph prompt learning adapts pre-trained GNNs for various tasks.
• Current methods prompt all graph elements, leading to redundancy and sub-optimality.

Purpose of the Study:

• To develop a more efficient graph prompting method by leveraging sparse representation theory.
• To introduce Graph Sparse Prompting (GSP) for adaptive and selective element prompting.

Main Methods:

• Proposed Graph Sparse Prompting (GSP) to sparsely select optimal graph elements.
• Introduced two GSP models: Graph Sparse Feature Prompting (GSFP) and Graph Sparse multi-Feature Prompting (GSmFP).
• Developed a simple yet effective algorithm for solving GSFP and GSmFP models.

Main Results:

• GSFP and GSmFP offer a general scheme for tuning pre-trained GNNs using sparsity-guided prompt learning.
• Experiments on 16 benchmark datasets demonstrated the effectiveness of the proposed GSFPs.
• The proposed methods achieve compact prompting by adaptively selecting desired attributes.

Conclusions:

• Graph Sparse Prompting (GSP) provides an efficient and effective method for adapting pre-trained GNNs.
• The proposed approach reduces redundancy by selectively prompting graph elements.
• The GSP framework shows significant advantages across various graph learning tasks.