Abstract:Traditional recommendation algorithms rely solely on users’ historical rating data to identify similar users, which will suffer from severe data sparsity and cold start problems and thus result in unsatisfactory recommendation performance. A recommendation algorithm model based on multi-source feature fusion (MSFF) is proposed to address data sparsity and improve recommendation quality, which deeply mines the semantic information in review texts and rating data. This model employs a hybrid model of CNN and SA_BiLSTM to extract features from review texts, and introduces a self-attention mechanism during feature extraction to dynamically adjust the importance of review information, thereby better capturing the deep key information in the review information. Subsequently, the latent factor model (LFM) is adopted to decompose the rating matrix to obtain latent features of users and items. Then, the non-homogeneous features obtained from the two models are effectively fused to generate high-order features, which are adopted to predict ratings. Finally, a recommendation list is generated based on the predicted ratings. Experiments are conducted on the Yelp review dataset and three sub-datasets in the Amazon review dataset, Amazon5-core, to validate the effectiveness of the model. The results demonstrate that the model not only mitigates the influence of rating data sparsity in recommendation algorithms but also improves the accuracy of recommendation results.