TRWH: A Text-Driven Random Walk Heterogeneous GNN for Semantic-Aware Sparse Recommendation

The paper introduces TRWH, a recommendation framework that combines Word2Vec and LLM-generated user and item profiles with a heterogeneous graph neural network. Random-walk augmentation adds second-order user-user and item-item links to sparse interaction graphs. On the Amazon-2023 Fashion and Beauty datasets, the authors report substantial RMSE and MAE reductions against state-of-the-art baselines. However, the abstract also reports an important trade-off: random walks help traditional embeddings but may dilute nuanced LLM representations, motivating adaptive integration strategies.