Model Based Collaborative Recommender System by Integrating Word Embedding and Matrix Factorization

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recommender systems assist users in identifying relevant products or services by predicting preferences from available interaction data. Collaborative filtering commonly relies on user-item ratings, while model-based approaches learn predictive representations from these ratings. Matrix factorisation is widely used in model-based collaborative filtering because it reduces dimensionality and can address sparse rating matrices. However, standard matrix-factorisation approaches generally do not incorporate the semantic information contained in user reviews. This study proposes a model-based collaborative recommender system that integrates user-review word embeddings with matrix factorisation. User reviews are converted into numerical representations using one-hot encoding and the Keras Embedding Layer, and the resulting embedding matrix is incorporated into a modified Singular Value Decomposition formulation, termed SVDEm. The proposed approach is evaluated against standard SVD and SVD++ using the Amazon Health and Hotel Reviews datasets under 5-fold cross-validation. Performance is assessed using RMSE, MAE, R-squared, and Explained Variance. SVDEm produces the lowest RMSE and MAE values on both datasets and achieves R-squared and Explained Variance values of 0.10 on each dataset, compared with lower corresponding values for the baseline approaches. The observed improvements are modest but consistent across the reported evaluation measures. These results indicate that incorporating semantic information from user reviews into matrix-factorisation-based collaborative filtering can improve rating-prediction performance relative to the evaluated standard SVD and SVD++ models.