Research On False Comment Detection Method Of Fusion Content And Behavior

With the development of the Internet, especially the rapid development of e-commerce, more and more consumers prefer shopping online, consumers are more likely to comment on their own products, and these product review information offers valuable information resources for manufacturers and potential consumers. Because there is some interest, some of which may be untrue or false content. To some extent, these fake reviews affected the reference value of review information, misleading consumers. So it is particularly important to detect fake reviews. The most basic review information is review content information, and mine it. It plays an extremely important role using review content information to detect fake reviews. In addition, mine reviewer behavior, it also plays an important role to detect fake reviews by finding abnormal behavior patterns. This paper focus on product and service reviews, which does research on the following key problems around detecting fake reviews based on review contents, detecting fake reviews based on reviewer behavior, and detecting fake reviews fuse these two features review content and reviewer behavior. Mainly completed the following research work:(1) A fake review detection method based on review content is proposed. At first, we construct topic-opposite sentiment dependency model(TOSDM) of reviews based on sentiment dependency, which is used to extract review topic information and their corresponding emotional information.Then analyze and extract6types of review content features combined review topics and emotional information. Finally, exploiting all these review content features which were extracted in the back step to detect fake reviews using classifier based on supervised method.(2) A fake review detection method based on reviewer behavior is put forward. Firstly, according to the reviewed data, we extract6types of features which represent reviewer behavior, and normalize the features of each dimension. Secondly, we build a clustering matrix based on the features of each review using F statistic to improve the K-means algorithm, and to achieve adaptive clustering for reviews. Finally, we calculate the degree of deviation from the entire review data set for each cluster, and determine abnormal clusters based on the threshold value to achieve fake review detection. (3) With the thought of semi-supervised learning, a fake review detection method based on review content and reviewer behavior is proposed. We first extract the features of review content and reviewer behavior, then with semi-supervised learning ideas for Co-Training, take these two types of features as a separate view, establish classifiers using of these two types of separate features, and then select unlabeled samples whose confidence is high. Finally, using these selected samples to update the training mode to improve the effect of the classifier.(4) Design and implement a prototype system to detect fake reviews, which could provide convenience for further study on the detection of fake reviews.