Temporal GNN-LLM Fusion

A fusion architecture that injects graph representations into a language model with PEFT to generate text associated with social-network edges.

Author: Xuhui Zhan

Project affiliation: Network and Data Science Lab at Vanderbilt University · Role: Graduate Research Assistant

The author line identifies who wrote this page. Xuhui Zhan is the primary contributor to the project, not necessarily its only contributor.

Advisor: Tyler Derr

What can a language model recover from relationships that text alone does not expose? We studied how graph-derived representations can condition language generation for social-network edges. The proposed fusion network injects a Graph Neural Network representation into the language model, while parameter-efficient fine-tuning (PEFT) adapts a limited set of parameters for the generation task.

The initial empirical evaluation used the Venmo Dataset. We planned an additional evaluation on the Amazon Review Dataset to test cross-domain transfer rather than assuming that results from one graph and text distribution would generalize to another.