Large Language Models for Graphs: Progresses and Directions

Graph neural networks (GNNs) have emerged as fundamental methods for handling structured graph data in various domains, including citation networks, molecule prediction, and recommender systems. They enable the learning of informative node or graph representations, which are crucial for tasks such as link prediction and node classification in the context of graphs. To achieve high-quality graph representation learning, certain essential factors come into play: clean labels, accurate graph structures, and sufficient initial node features. However, real-world graph data often suffer from noise and sparse labels, while different datasets have unique feature constructions. These factors significantly impact the generalization capabilities of graph neural networks, particularly when faced with unseen tasks. Recently, due to the efficent text processing and task generalization capability of large language models (LLMs), there has been a promising approach to address the challenges mentioned above by combining large language models with graph data.

This tutorial offers an overview of incorporating large language models into the graph domain, accompanied by practical examples. The methods are categorized into three dimensions: utilizing LLMs as augmenters, predictors, and agents for graph learning tasks. We will delve into the current progress and future directions within this field. By introducing this emerging topic, our aim is to enhance the audience's understanding of LLM-based graph learning techniques, foster idea exchange, and encourage discussions that drive continuous advancements in this domain.