Multi-criteria Recommender System Based On Text Mining

Given the development and popularization of the Internet,the amount of available information has increased dramatically,thereby resulting in the problem of information overload.A personalized recommendation system is one effective way to solve this problem and has been used in many applications.However,in the e-commerce environment,obtaining sufficient rating information is difficult,thereby giving rise to the problem of data sparsity.To address this issue,rich user-generated contents should be incorporated into the process of user modeling and recommendation generation.Two review-based multi-criteria recommendation methods were proposed in this paper: the fuzzy sentiment rating-inferred(FSRI)method and the fuzzy sentiment comment-direct(FSCD)method.FSRI uses emotional intensity extracted from usergenerated reviews to replace the numerical ratings used in traditional memory-based collaborative filtering recommendation.FSCD characterize users and items by the textual content of their reviews directly.Both methods are based on our proposed fuzzy sentiment strength detection approach,which utilizes the HowNet lexicon to give words different emotional intensities.To obtain fine-grained aspect preferences,we propose a semi-supervised framework for online review classification using information gain(IG)for feature selection and support vector machine(SVM)to determine the criterion category to which every sentence belongs.Moreover,we extend the similarity between users or users and items to the multi-criteria recommendation field.Our approach has two main advantages over the existing review-based recommendation methods.First,we emphasize user multi-criteria preferences depicted in their reviews.Second,we use fuzzy logic methods to address the imprecision and vagueness of the natural linguistic descriptors.Our experiment results demonstrate the excellent performance of our proposed two methods and show that our review-based recommendation approach that uses textual information directly performed better than our rating-inferred approach in most cases.