TRWH: Text-Driven Random Walk Heterogeneous GNN for Sparse Recommendation
A research paper introduces TRWH (Text-driven Random Walk Heterogeneous Graph Neural Network), a framework that combines Large Language Models (LLMs) with heterogeneous graph structures to improve recommendation systems in sparse data settings. TRWH uses LLM-generated textual profiles and random walk augmentation to create user and item representations. It has three components: Embedding Creation using Word2Vec and LLM profiling, a Heterogeneous Graph Neural Network (HeteroGNN) for multi-relational information propagation, and Random Walk-based Path Construction to add second-order user-user and item-item links. Experiments were conducted on the Amazon-20 dataset. The paper is available on arXiv under ID 2607.25471.
Key facts
- TRWH stands for Text-driven Random Walk Heterogeneous Graph Neural Network.
- The framework fuses LLM-generated textual profiles with heterogeneous graph structures.
- It uses random walk augmentation to address sparse recommendation settings.
- Three core components: Embedding Creation, HeteroGNN, and Random Walk-based Path Construction.
- Embedding Creation uses both Word2Vec and LLM-based profiling.
- HeteroGNN propagates information across multi-relational edges.
- Random Walk-based Path Construction enriches sparse graphs with second-order links.
- Experiments were performed on the Amazon-20 dataset.
Entities
Institutions
- arXiv