CF ‐ SBERTHet : Collaborative and Textual Knowledge Enhanced Semantic Graphs for Sparse Recommendations

ABSTRACT Modern e‐commerce platforms face a critical challenge: delivering accurate recommendations under extreme user–item interaction sparsity, where textual context remains systematically underutilised. Existing collaborative filtering methods degrade sharply in sparse settings, while semantic approaches fail to capture collaborative patterns effectively. We propose CF‐SBERTHet, a unified recommendation framework that integrates collaborative filtering signals with rich semantic information extracted from user reviews through heterogeneous graph neural networks. Our approach generates enriched item embeddings by aligning representations from a pre‐trained collaborative filtering model with semantic features extracted via Sentence‐BERT, and then models complex multi‐relational interactions through type‐specific message passing over seven distinct edge types, encompassing ratings, reviews, purchases, and contextual item similarities. Extensive experiments on four Amazon 2023 datasets (Fashion, Beauty, Musical Instruments and Movies and TV) demonstrate that CF‐SBERTHet consistently outperforms competitive baselines across all evaluation settings in terms of both Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), achieving robust performance even at interaction sparsities exceeding 99%. Comprehensive ablation studies confirm that each component, namely the collaborative and textual knowledge‐enriched embeddings, the heterogeneous graph architecture, and the review‐based edge weights, contributes critically to overall performance. These results establish CF‐SBERTHet as an effective and broadly applicable solution for context‐aware recommendation in sparse real‐world e‐commerce environments.