Project

Knowledge Injection for Text Generation in Social Network

A fusion network that combines Graph Neural Network representations with LLMs through knowledge injection and PEFT to generate textual edge features in social networks.

Illustration of the knowledge injection framework

Advisor: Tyler Derr

I explored how graph structure can inform language generation. A fusion network injects representations from a Graph Neural Network into an LLM, while parameter-efficient fine-tuning (PEFT) adapts the language model to generate text-related edge features for social networks.

The initial empirical evaluation used the Venmo Dataset. Additional experiments were planned for the Amazon Review Dataset to test how well the approach transfers across graph domains.

↵