Recommendation Algorithms Based on User Reviews and Scoring
Yueyi Wang, Xiaoyan Zhang · 2023
This study offers a recommendation algorithm based on user reviews and ratings to overcome the sparsity issue of the rating matrix and the problem of incorrect suggestions resulting from utilizing rating data alone. Two parallel convolutional neural network models are used to extract the text features in user and item reviews simultaneously by setting up shared layers to obtain two matrices of user preferences and item attributes about the reviews, which are then combined to generate the rating matrix for review prediction. The rating matrix is dissected using a hidden semantic model approach to create two matrices of user preferences and item characteristics about ratings. These are then merged to obtain the rating matrix for rating prediction. For the feature fusion portion of reviews and ratings, the predicted rating matrices for reviews and ratings are combined using a multilayer perceptron architecture to generate the final rating prediction matrix. The algorithmic model suggested in this research fuses review and rating data and is thus able to rectify inaccurate data during feature fusion and has some improvement on the cold start issue, when fresh users or objects are available with inadequate data for modeling. Experiments demonstrate that the algorithmic model suggested in this research is superior to the original recommended algorithmic model for ratings or reviews regarding rating prediction accuracy.