Design And Implementation Of Recommendation Algorithm Based On Review Analysis

With the rapid growth of Web2.0, it provides users an opportunity to interact with each other. One of the representative application scenario is product comment platform. Users are allowed to write reviews including a piece of description content and an overall evaluation score after the experience of the service. Review content reflects users’satisfac-tion and opinions on products they bought, and it helps potential consumers make choice. Such kind of user generated content is exploding and brings new opportunity for product organization, recommendation, user profiling and so on.Recommender systems are widely deployed in Web applications that need to predict the preferences of users to items. They are popular in helping users find movies, books and products in general. Analyzing user ratings and review content can discover item profile that users focus on. Based on item profile, it can predict how much the user would like the item. We consider designing recommendation algorithms based on review content analysis. On the other hand, most users have different backgrounds or preferences which make reviews/ratings different. We also propose a method to evaluate user credibility on aspects.Main contributions are summarized as follows:· An Overall Rating Prediction Based on Modeling Profiles We design a method for item recommendation based on a novel model that captures correlations between hidden topics in reviews and numeric ratings. It is motivated by the observation that a user’s preference against an item is affected by different topics discussed in reviews. Our method first explores topic modeling to discover hidden topic distri- bution from review text. Profiles are then created for users and items separately based on topics discovered in their reviews. Finally, we utilize regression models to detect the user-item relationship and the rating is modeled as the similarity between user and item profiles.· Multi-dimensional Rating Prediction Based on Semantic Analysis We propose a method to predict user ratings on aspects considering different customers caring about different things. We first use a bootstrapping algorithm to detect aspects from review texts. Then, it utilizes the polarity of emotional words to calculate the rating on each aspect.· Evaluation of User Credibility on Aspects We explore user credibility on multi-dimensions by considering users difference on aspects. An iteration algorithm is designed for estimating credibility by considering the consistency between individ-ual ratings and overall ratings on aspects. Here, we assume that an overall rating from a user is the weighted combination of ratings corresponding to different aspects in review text.