Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

Towards Text-free Graph Foundation Models: Rethinking Multi-Domain Graph Contrastive Learning

Created by
  • Haebom

Author

Zihao Zhao, Xinlong Zhai, Jinyu Yang, Chuan Shi

Outline

In this paper, we propose MDGCL, a novel multi-domain pre-learning and cross-domain transfer framework that builds graph-based foundation models by leveraging graph data from various domains. To overcome the limitations of conventional single-domain-centric contrastive learning strategies, we introduce a contrastive learning strategy that recognizes and captures differences between domains and domain tokens that encode domain-level global information. In downstream tasks, we enable fine-grained domain knowledge transfer through a domain attention mechanism. Experimental results on five benchmark datasets demonstrate that the proposed method outperforms state-of-the-art techniques by up to 19.33% in accuracy and up to 19.13% in Macro-F1 score.

Takeaways, Limitations

•
Takeaways:
◦
We present a new framework that overcomes the limitations of existing single-domain-centered graph-based basic model learning methods and effectively utilizes multi-domain graph data.
◦
We propose a contrastive learning strategy and a domain attention mechanism that integrate knowledge from various domains by effectively considering differences between domains.
◦
Achieves state-of-the-art performance on a variety of graph-based tasks, demonstrating substantial performance improvements.
•
Limitations:
◦
The effectiveness of the proposed method is based on experimental results on specific benchmark datasets, and generalization performance to other types of graph data or tasks requires further study.
◦
The design of domain tokens and domain attention mechanisms is limited to a specific method, and there is a lack of comparative analysis on the effectiveness of other design methods.
◦
There may still be challenges in obtaining and preprocessing graph data from various domains.
👍
Made with Slashpage