Deep recommendation algorithm based on reviews and descriptions neural matrix factorization under cold-start

Existing recommendation methods rely solely on ratings or review interactions to determine user preferences, resulting in inability to make recommendations during the cold-start phase of a new item. This paper proposes a RDNMF(Reviews and Descriptions Neural Matrix Factorization) model that integrates review text, item description documents, and ratings. First, the pre-trained BERT model provided by Google is used to process review text and item description documents. LSTM is trained to analyze changes in user preferences over time. The impact of review text on user preferences is then analyzed. Next, user latent feature vectors and item latent feature vectors, both based on review text and item descriptions, are deeply mined. Finally, latent feature vectors obtained by decomposing the rating matrix, are input into an improved neural matrix factorization model. Experiments were conducted on four publicly available Amazon datasets and the BookCrossing dataset, and compared with various state-of-the-art algorithms. The results indicate that the recommendation performance of RDNMF has been improved. Among which, HR@10 improved by an average of 3.43%, and the RMSE decreased by an average of 2.78%, NDCG@10 improved by an average of 5.85%.