Research On Matrix Decomposition Recommendation Algorithm Integrating Sentiment Analysi

Traditional recommendation algorithms directly use user’s rating for prediction,which usually suffer from rating noise.The noise results from user’s random rating behavior and providers maliciously promoting their products through fake ratings.Sentiment analysis on reviews can effectively denoise ratings and obtain fine-grained ratings.Aiming at the above problems,this thesis studies the most popular matrix factorization algorithm,and introduces sentiment analysis and reliability to improve the recommendation accuracy.The main research contents include:First,this thesis proposes a rating-based sentiment dictionary construction method,which can extract sentiment words with corresponding sentiment from reviews based on ratings.Compared with traditional dictionary,each word in the sentiment dictionary not only has a sentiment label(positive/negative),but also a score of its sentiment level.In addition,this thesis constructs sentiment dictionaries on different datasets,which conform to the text characteristics of reviews,and alleviate the phenomenon that the same word have different sentiment tendency under different text environment.Secondly,in terms of leveraging reviews to assist ratings,a sentiment-based matrix factorization(SBMF)algorithm framework is proposed.Based on the sentiment dictionary,the sentiment scores of reviews are obtained and integrated into the matrix factorization framework.Unlike simple summation,SBMF enables the implicit feature matrix of users/items to map users’ ratings and reviews’ sentiment features,which can effectively leverage reviews to assist ratings.Finally,in terms of the inconsistency between reviews and ratings,this thesis proposes a sentiment-based matrix factorization with reliability(SBMF+R)algorithm.On one hand,by calculating the Euclidean distance between user’s rating and sentiment vector,the consistency of user’s review and rating can be measured.On the other hand,review’s helpfulness votes can reflect its’ quality and reliability.We use credibility as the weight of ratings and reviews to integrate into the framework of SBMF and prediction.Experiments are conducted on eight Amazon datasets and compared with state-of-the-art algorithms.The experimental results demonstrated that SBMF+R has achieved progress in terms of prediction accuracy.