Research On Review-based Rating Prediction Method For Recommendation

With the improvement of Internet infrastructure and the maturity of cloud computing,information technology has been widely used in all walks of life,which not only improves people's living standards and work efficiency,but also makes exponential growth in the valuable network data.It becomes more and more difficult to efficiently retrieve the required information when faced with such large scale data,which is the so-called “information overload” problem.By analyzing the historical behavior data of the user,the recommender system can actively extract information of interest to users from the large-scale data,it has become one of the effective ways to alleviate “information overload”.However,there is still a large amount of challenges:(1)Data Sparseness Problem: user and item interaction records are usually stored in a matrix especially in a system with a large number of items.However,a user often only interacts with a small number of items and inevitably leads to a relatively sparse row of the matrix,which increases the difficulty of modeling the user preference.(2)Cold Start Problem: when a new user(or item)joins the system,a new row(or column)is accordingly added to the interaction record matrix.In this case,the recommendation service is unsatisfactory due to the data used to model the new user(or item)is insufficient.(3)Interpretability Problem: the present most recommender systems only provide users with model prediction results without the recommended reasons,which greatly affects the user experience.Therefore,how to combine the recommended reason with the recommendation result is a question worth exploring.In view of the above problems,this paper combines the deep learning technology of natural language processing to conduct research on review-based recommendation algorithms.The key contributions are as follows:(1)Aiming at the problem of sparse data and cold start,a recommendation model combing review text with rating matrix is proposed.In recent years,related studies have only used rating data or review data.In order to use both of them,the pre-trained BERT model isused to provide the general semantic information of reviews and the adopted attention mechanism measures the contribution of each review to obtain user and item characteristics.Second,to better integrate the review data and rating data into a whole deep learning neural network model,the traditional factorization machine model is equivalently transform into its neural network form.The experimental results show the model can achieve a better performance of rating prediction.(2)For the interpretability problem,the multi-level fine-grained interaction model is proposed.Since the model mentioned above and other recent works ignore the issue of dynamic encoding(i.e.,user preferences should change with different items),four interactive attention mechanisms are introduced.Then,considering the issue that only provide one side of the explanation in related works,the attention mechanism of review level and aspect level are designed to improve the interpretability of recommendation results from multiaspect.Experimental results show that the model can provide interpretable results at the review level and the aspect level,respectively.