Aiming at the problem that the sparse rating data leads to the degradation of the recommendation quality of collaborative filtering algorithm, a multi-source deep recommendation model of fusion dual attention mechanism (MSDA) is proposed by fully mining reviews to enhance the recommendation performance. The model is based on three information sources of rating data, user review sets and product review sets for recommendation. It combines convolutional neural network and dual attention mechanism to mine the text features of reviews, and uses neural factorization machine to perform high-order nonlinear interaction between rating and review features, so as to achieve rating prediction. Experimental results show that, compared with advanced benchmark models such as NeuMF, NARRE, HRDR, MSDA significantly improves the score prediction performance.

data sparsity; review text; deep learning; rating prediction