A research achievement by Dr. Songwei Zhao of the School of Artificial Intelligence on graph neural network aggregation has been accepted byArtificial Intelligence, a premier international journal in the field of artificial intelligence. It is the first research result published by the School in this journal. Founded in 1970,Artificial Intelligenceis one of the longest-established and most influential authoritative academic journals in artificial intelligence. It is a Category A journal in artificial intelligence recommended by the China Computer Federation (CCF), and publishes original research results in areas including cognition and artificial intelligence, knowledge representation, machine learning, planning and action, and intelligent decision-making.
Towards Context-Aware Graph Representation Learning: Adaptive Node Aggregation with LLMs
First author: Songwei Zhao
Journal:Artificial Intelligence
Corresponding authors: Hechang Chen, Yuan Jiang
Overview
Representation learning on graph-structured data is an important research area in graph machine learning, with broad applications in real-world settings such as social networks and recommender systems. Conventional graph neural networks (GNNs) typically rely on static graph aggregation functions (e.g., mean, sum) applied to shallow text features, which limits their ability to model complex semantic interactions among nodes. A recent line of progress introduces large language models (LLMs) to enhance node features through deep semantic encoding. Most existing work, however, treats LLMs as static encoders, overlooking their potential to reason about inter-node relationships. In this work, we propose LLM-based Tailored Node Aggregation (LTNA), a novel approach that exploits the contextual reasoning capabilities of LLMs to generate adaptive, semantics-driven aggregation strategies for each node. Rather than relying on fixed rules or global functions, LTNA prompts the LLM to infer the relational meaning between a node and its neighbors, and outputs a tailored aggregation strategy based on the inferred semantics. This allows message passing in GNNs to go beyond numeric aggregation and become a context-aware semantic operation. By jointly exploiting node textual attributes and graph structure, LTNA enhances the expressiveness of node representations and improves performance across downstream tasks.
