Research On Recommendation Algorithm Based On Reviews Of Users

With the rapid development of the Internet,information data has badly exploded.In order to mine users' interest from massive data,recommendation algorithms have emerged and are widely used in all aspects of people's daily lives.The user-based collaborative filtering recommendation algorithm is one of the most commonly used recommendation algorithms.However,the traditional collaborative filtering algorithm also faces serious data sparsity and cold start problem.When there are few ratings of products or users,rating prediction and product recommendation aren't effective at all.The review text contains a wealth of product information and user opinions,which can be used to improve the system's recommendation accuracy.Thus,it has greatly attracted the scholars' interest in recent years.In this paper,we try to use the review text and rating to improve the neighbor-based collaborative filtering recommendation algorithm.Our contributions are as follows:(1)The characteristics of review text are deeply analyzed.When traditional TFIDF algorithm is applied to short review text,there's little difference in word frequency distribution between significant terms and secondary terms.Therefore,key information in the text cannot be extracted effectively.The helpful feedback of a review is introduced to the TFIDF algorithm to distinguish significant vocabularies and secondary ones,so key information can be extracted effectively from the review.(2)When extracting keywords,the traditional TextRank algorithm doesn't take the significance of different words into consideration and divides the weight equally when an origin lexical node jumps to the adjacent nodes.In view of this,we combine TextRank algorithm with the improved TFIDF algorithm to reboot the weight assignment process,which enhances accuracy of keyword extraction.However,the keywords extracted by the improved TextRank algorithm still contain many synonymous and repeated words.The subsequent calculation for user preference similarity will be greatly affected if dimensionality reduction progress is not executed.In this paper,singular value decomposition algorithm is used to map product-keyword matrix into latent semantic space and classify keywords based on their semantics so as to discover the feature of product topic distribution.(3)The traditional collaborative filtering recommendation model fairly relies on user ratings.To solve this problem,we introduce topic distribution feature of products to the rating prediction model.The topic preference similarity between every two users is calculated by counting the topic distribution of products that scored by them.The overall users' similarity is then fused with the topic preference similarity and rating similarity.As a result,the error for rating prediction can be reduced.(4)Our proposed algorithm is verified on Amazon's movie&TV dataset and compared with three baseline algorithms.The experiment proves that our proposed algorithm can improve the accuracy of rating prediction and alleviate the data sparsity problem to some extent.