Research On Hybrid Recommendation Algorithms Based On Rating Matrix And Review Texts

With the rapid development of the Internet,users are facing the problem of information explosion.The recommendation system assists users in making decisions by providing personalized product recommendation services,thereby alleviating consumers' anxiety and confusion caused by excessive choices.The algorithm based on the rating matrix can explicitly model users and items according to the user's rating patterns,but the algorithm faces the problem of data sparsity.The user-generated review text can provide rich semantic information of user preferences and item features,which can effectively alleviate the problem of data sparsity and improve the accuracy of model recommendation.This paper is mainly based on deep learning technology to study a hybrid algorithm that combines a rating matrix and review text.The main research results obtained are as follows:(1)A text analysis model of bidirectional GRU based on Bert and attention mechanism is proposed.Firstly,BERT is introduced as a word embedding layer to extract the semantic information of the review text.Secondly,because the CNN convolution kernel has a fixed size and can not effectively deal with the text sequence,a bidirectional GRU is proposed to encode the review text and capture the context information of the word from both directions.So the model can understand more accurately the meaning expressed in the review text.Finally,by introducing an attention mechanism to assign different weights to each review,the model can focus on more useful reviews.Simultaneously,the deep matrix decomposition model is used to extract scoring features,which solves the problem that the traditional matrix decomposition model can only extract low-order and linear features,and enhances the ability of interactive learning between features vectors.(2)A deep fusion method based on rating information and review text features is proposed.Since the latent features extracted based on the rating matrix and the review text have different origins,they cannot be jointly learned after being output to different feature spaces.The model proposed in this paper realizes the interactive learning between multi-source features by cascading factor decomposition machine and multi-layer perceptron,and further improves recommendation accuracy.Finally,the proposed model is experimentally verified on five data sets provided by Amazon.The experimental results show that the model proposed in this paper has the lowest score prediction error rate.Compared with several excellent deep recommendation benchmark models based on comment text,the average prediction errors MSE and RMSE are reduced by about 3.99% and 2.04%,respectively.