Research On Personalized Recommendation Algorithms Fusing User Reviews And Ratings

With the rapid growth of e-commerce shopping website users and goods,the needs of users and the types of goods are also changing,personalized recommendation system began to face the problem of data sparsity.In the personalized recommendation system,the traditional recommendation algorithm in the face of sparse rating data,due to the lack of common rating information,the recommendation results are difficult to be accurate.Therefore,it is very important to solve the sparsity problem of rating data.And fusion recommendation is an effective method to solve the problem of score sparsity.Fusion recommendation is to fuse the rating data and other types of user feedback data,such as comment text,image,etc.,into the traditional recommendation algorithm,and use the rich user preference information and product feature information of user feedback data to fill in the sparsity of rating data.Some scholars have improved the basic matrix decomposition model and BP neural network model to achieve the fusion recommendation algorithm.Most of the existing fusion recommendation algorithms based on matrix factorization are limited by the effectiveness of simple fusion operators to extract deep features,which easily leads to the lack of information,resulting in poor fusion effect,and manual design also consumes more time and energy.Similarly,the existing fusion recommendation algorithm based on BP neural network still needs the intervention and assistance of traditional recommendation algorithm to complete the task of fusion recommendation.The fusion process is more complex,and it depends on the connection between neural network and other algorithms,and its execution efficiency is low.It can be seen that the current two fusion recommendation algorithms still have limitations.It is necessary to further reduce the complexity of their fusion,improve the fusion effect,and solve the problem of data sparsity more efficiently.Therefore,this paper starts from two kinds of improvement ideas,namely the idea of improving learning objectives and the idea of improving the structure of the model,and makes corresponding improvements on them.Based on the idea of improving learning objectives,an improved matrix factorization fusing reviews(MFFR)is proposed.The learning goal of MFFR is no longer a single rating data,but learning rating data and comment text at the same time.The two kinds of data are directly fused in the process of resolution.On the other hand,an improved neural network fusing recommendation model(NNFR)is proposed based on the improved model structure.The fusion layer is designed in the network,and the two types of data are directly fused in the fusion layer,which avoids the intervention of traditional recommendation algorithms in previous studies.The experimental results on three public data sets of e-commerce shopping websites show that,compared with the traditional recommendation algorithm,the two fusion recommendation models proposed in this paper have better performance,which can effectively mine the deep-seated characteristics of users,so as to fill in the sparse information of the original rating data and obtain higher recommendation accuracy.