Aligning Graph Models with Large Language Models for Open-Task Scenarios

Speaker: Yang Cheng, Associate Professor at Beijing University of Posts and Telecommunications, specializes in data mining and natural language processing, with over 30 CCF A‑class papers and more than 9,000 citations, and has received the 2020 Ministry of Education Natural Science Award (4th prize) among other honors.

Talk Title: Aligning Graph Models with Large Language Models for Open-Task Scenarios

Outline: Large language models such as ChatGPT demonstrate strong zero‑shot learning and instruction‑following abilities for a variety of open tasks described in natural language. However, graph‑structured data analysis still relies on graph neural networks that are limited to predefined tasks like node classification or link prediction. To address this gap, we propose GraphTranslator, which connects pretrained graph models with large language models. The graph model handles predefined tasks, while the language model serves as an extensible interface for open‑ended tasks. To train GraphTranslator, we design a Producer that automatically constructs node‑text alignment data, including node information, neighbor context, and model metadata. By translating node embeddings into tokens, GraphTranslator endows the language model with the capability to make predictions based on linguistic instructions, offering a unified solution for both predefined and open tasks.

Audience Benefits: Attendees will learn how to leverage large language models for graph data analysis, understand methods for integrating graph neural networks with language models, and explore practical techniques for building alignment datasets that enable open‑task capabilities.

Scan the QR code to register for free.