A Study On User Reputation And UGC Quality Evaluation Model

Web2.0has promoted the prosperity of the user generated content which not only promotes the dissemination of knowledge but also brings about problems such as damag-ing UGC content of others, making up fake reviews and uploading result of others without respecting copyright. All these problems make it necessary to evaluate user reputation and user generated content quality.This article first elaborated the data acquisition, preprocessing and feature extraction. We took the English Wikipedia as a representative of the knowledge sharing sites for research. We downloaded the English Wikipedia data and extracted a large number of features by means of the characteristics of editors and articles in Wikipedia. We selected three categories as our research object, and labeled editors’reputation and articles’quality under these three categories to get a golden-standard dataset. We took the Amazon website as a representative of E-commerce sites for research. We downloaded the Amazon review dataset, selected the reviews for electronic product as our research object. Firstly, to solve the sparseness in dataset we preprocessed our dataset and extracted a large number of features by means of the characteristics of reviewers and reviews in Amazon. Then we labeled reviewers’reputation and reviews’quality to obtain a golden-standard dataset.In Wikipedia, we proposed a dual wing factor graph model which utilizes features we have defined to unify the tow difficult problems, assessment of articles’quality and editors’reputation into one united model. Then we used the L-BFGS algorithm to learn our model. Moreover, we conducted experiments on test data to validate our method and compared results with several baseline methods. Experiments showed that our approach obtains significant improvement over these baseline methods on accuracy and F1mea-surement. In amazon, we proposed a review factor graph model. This model incorporates all features we have extracted to detect fake reviewers and fake reviews. We utilized the L-BFGS algorithm to learn our model. Then we conduct experiments on test data to vali-date our method and compare results with other baseline methods. Finally, we summarize our work and put forward the focus of the future research direction.