This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
This paper explores the potential of large-scale language models (LLMs) for Continual Learning (CL) problems that process graph-structured data. Specifically, we address challenges in the experimental setups of existing Graph Continuous Learning (GCL) research and evaluate the performance of LLMs in more realistic scenarios. Our results demonstrate that LLMs can achieve superior performance with only minor modifications. Building on this, we propose Simple Graph Continual Learning (SimGCL), a method that outperforms existing state-of-the-art GNN-based techniques by approximately 20% without rehearsal. Furthermore, we develop the LLM4GCL benchmark to enhance the reproducibility of GCL methods.
Takeaways, Limitations
•
Takeaways:
◦
We propose the possibility of alleviating the catastrophic forgetting problem in graph continuous learning by utilizing LLM.
◦
Point out problems arising from the experimental setup of GCL research and suggest a more realistic evaluation method.
◦
We propose a simple and effective method called SimGCL, which achieves performance that surpasses existing SOTA.
◦
Contribute to the advancement of research by developing the LLM4GCL benchmark to increase the reproducibility of GCL research.
•
Limitations:
◦
The paper may lack a detailed description of how to use LLM specifically, i.e., the detailed implementation of SimGCL.
◦
There are parts that depend on the performance of LLM, so the performance of SimGCL may be affected depending on the development speed of LLM.
◦
Further research may be needed to determine the generalization ability of the proposed method.
◦
It may only be effective for certain GCL tasks, and performance verification for other GCL tasks is needed.
