Multilevel Context Learning with Large Language Models for Text-Attributed Graphs on Social Networks.

Area of Science:

• Social network analysis
• Graph representation learning
• Natural Language Processing (NLP)

Background:

• Social networks contain complex graph structures and rich textual information.
• Existing text-attributed graph (TAG) methods often fail to capture multilevel context, leading to information loss.
• Current frameworks convey graph structures to large language models (LLMs) for semantic embeddings but lack comprehensive context.

Purpose of the Study:

• To introduce a novel model, the Multilevel Context Learner (MCL), for enhancing LLMs' semantic embedding capabilities on social networks.
• To address the limitations of existing methods in capturing multilevel context within social network data.
• To improve the representation learning for downstream graph neural network (GNN) tasks.

Main Methods:

• Modeling social networks as a multilevel context textual-edge graph (MC-TEG) to capture both structure and semantic relationships.
• Leveraging LLMs' reasoning capabilities to generate semantic embeddings by integrating multilevel contexts.
• Employing tailored bidirectional dynamic graph attention layers to refine weight information.

Main Results:

• The MCL model consistently outperformed all baseline models across six real-world social network datasets.
• Achieved significant prediction accuracy improvements, with absolute gains ranging from 6.11% to 11.05% over the next best models.
• Demonstrated superior performance in generating semantic embeddings by effectively utilizing multilevel contextual information.

Conclusions:

• The proposed MCL model is effective in enhancing semantic embeddings for social network analysis.
• Integrating multilevel context is crucial for improving representation learning in text-attributed graphs.
• The MCL model offers a promising approach for various downstream GNN tasks on social networks.