Research On Deep Recommendation Algorithm Combining Review Information

With the improvement of Internet technology,the way people obtain and publish information is becoming more and more convenient,and the amount of data of various kinds of information is also increasing.At the same time,there is the problem of "information overload",where people cannot efficiently find the information they want from a large amount of data.Recommendation systems are used to solve this problem.In the face of vague user needs,the recommendation system calculates the user's preferences based on the user's historical information and attributes,filters the things they may be interested in from a large amount of data,and then makes recommendations.In the face of constantly adding new users and new items,traditional recommendation algorithms cannot improve the accuracy of recommendation due to lack of historical data,which is the challenge brought by cold start.As a result,a lot of research work has emerged to improve recommendation performance by adding auxiliary information,such as user and item attributes,images,social relationships,etc.The review text is the most direct feedback from the user to the item,and often contains information such as the user's preference and the characteristics of the item.Learning user and item representations from review sentences can not only alleviate the problem of sparse rating matrices,but also improve the interpretability of recommendations.In view of how to use deep learning technology to more effectively use review information for rating prediction,this paper mainly does the following work:A recommendation algorithm that fuses review features and rating features is proposed.Extracting review features using convolutional neural networks.Using word-level attention,calculate the importance of words.Many related studies ignore the correlation between users and items,so interactive attention is added to calculate the specific correlation between users and items.In practice,many users do not have the habit of commenting,and short reviews do not contain enough information.To alleviate the problem of sparse review data,features are extracted from reviews of similar users as auxiliary review features of the current user to enhance the user representation.For the scoring feature,it is obtained through the user's and item's ID information.Finally,the different features are fused and fed into the factorization machine to output the predicted score.In similar reviews,the user's focus on items may not be completely similar,so there will be situations where the reviews are similar but the ratings are quite different.Therefore,we continue to improve the algorithm and propose a deep recommendation algorithm integrating personalized attention.Through the ID information used to uniquely identify users and items,personalized attention is generated,and the importance of similar reviews to different users is calculated.Finally,the auxiliary review features and user and item review features with personalized features are integrated to complete the scoring prediction task.Experiments are done on the Amazon data set.The evaluation standard selects the mean square error MSE value.By comparing experiments with different algorithms,it is proved that the two proposed algorithms have better performance.In addition,the validity of each part of the algorithm model is verified by experiments,and the effects of different parameter changes are analyzed in the experiments.