Item Rating Prediction Based On Machine Learning

In the era of rapid development of Internet technology and fast dissemination of information.Recommendation system has become an important approach to help users obtain effective information.Limited by its sparseness,the traditional recommendation algorithm cannot produce effective recommendation results in the case of a large scale of users and items.By using the user-item rating matrix,the matrix factorization model and its improved model can effectively solve the problems of high sparsity and low recommendation performance of collaborative filtering algorithm.In addition to the user's direct rating of the item,the review written by users can also represent user's preference and product's characteristic,thus many researchers proposed numerous rating prediction methods based on text feature extraction to alleviate the sparseness of the rating matrix.However,the above-mentioned methods singly use rating matrix or review text to predict and fail to combine them for comprehensive consideration,therefore the improvement of prediction precision is limited.To further improve the precision of rating prediction and the quality of recommendation,this thesis proposes a fusion model that combined the user-item rating matrix and review texts.The proposed model consists of three main modules,the matrix factorization and deep matrix factorization modules can extract the linear and non-linear interactions of users and items,while the convolutional neural network module can extract the semantic expression of users and items from rating texts in deep level,then the three modules were merged in the full connected layer to produce the final rating prediction result.At the end of the thesis,the proposed model is compared with other rating prediction models on Amazon's five rating and review datasets.The parameters that affect the performance of the model are analyzed and selected.The mean square error,accuracy,precision,recall and F1 value are selected to evaluate the prediction precision and the recommendation effect.The experimental results show that the prediction mean-square error of the Deep-Fusion score prediction model on different test sets is lower than other comparison models,and the recommendation effect is better.