# Can LLMs Convert Graphs to Text-Attributed Graphs?

**Authors:** Zehong Wang, Sidney Liu, Zheyuan Zhang, Tianyi Ma, Chuxu Zhang, Yanfang Ye  
**arXiv:** [2412.10136](https://arxiv.org/abs/2412.10136)  
**Date:** 2024-12-13  
**Summary page:** https://www.summarizepaper.com/en/paper/2412.10136v1/

## Key points

- Graph neural networks (GNNs) are essential for learning node embeddings in data analysis and machine learning applications.
- Existing GNN architectures face challenges when dealing with multiple graphs with differing feature spaces.
- Text-attributed graphs have been introduced to address the issue of cross-graph feature alignment by associating each node with a textual description.
- The Text-Attributed Network Synthesis (TANS) approach leverages large language models to convert existing graphs into text-attributed graphs, enhancing the understanding of how graph topology influences node semantics.
- TANS demonstrates superior performance on text-free graphs compared to manual feature design approaches, showcasing the potential of large language models for preprocessing graph data.

## Summary

, , , , 

In the realm of data analysis and machine learning, graphs serve as fundamental data structures that are prevalent in various real-world applications such as drug discovery, recommender systems, and social network analysis. Graph neural networks (GNNs) have emerged as a powerful tool for learning node embeddings by facilitating message passing on these intricate structures. However, a significant challenge arises when applying GNNs to multiple graphs with differing feature spaces, as existing GNN architectures lack provisions for cross-graph feature alignment. To tackle this issue, recent advancements have introduced the concept of text-attributed graphs where each node is associated with a textual description. This innovation enables the utilization of a shared textual encoder to project nodes from diverse graphs into a unified feature space. While this approach shows promise, it heavily relies on the availability of text-attributed data which can be challenging to acquire in practical scenarios. has been proposed in response to this gap in research. TANS leverages large language models (LLMs) to automatically convert existing graphs into text-attributed graphs. The core idea behind TANS is to integrate topological information with each node's properties, thereby enhancing the LLMs' capability to elucidate how graph topology influences node semantics. The efficacy of TANS was evaluated across text-rich, text-limited, and text-free graphs, demonstrating its ability to enable a single GNN to operate seamlessly across diverse graph structures. Particularly noteworthy is TANS's superior performance on text-free graphs compared to existing approaches that manually design node features. This showcases the immense potential of LLMs for preprocessing graph-structured data even in scenarios where textual information is absent. In conclusion, this research sheds light on an innovative approach that bridges the gap between graph neural networks and diverse graph structures through automated synthesis of text-attributed information using large language models. The code and data related to this research are available at https://github.com/Zehong-Wang/TANS.