Research On Detection Of User Anomaly Review Based On Deep Neural Network

With the rapid development of intelligent terminal and e-commerce,the impact of user's review information on people's daily life is becoming more and more important.People's dependence on review information catalyzes the emergence of abnormal reviews.Driven by economic interests,some people make huge profits by writing abnormal reviews,and even there is a black industrial chain that specially writes abnormal comments for others,which seriously affects people's shopping experience.In order to ensure users' comfortable shopping experience,many researchers have done a lot of exploration work on the abnormal review detection task of e-commerce websites.However,there are still many problems in the current model of anomaly review detection,such as it is difficult to capture the global semantics of the review text and there is no distinction between the key features of user's behavior,which leads to the low accuracy of the model detection,and the effectiveness of the user anomaly review detection task needs to be strengthened.In order to solve the above problems,this thesis studies the abnormal review detection as follows:1.Aiming at the problem that it is difficult to capture the global semantics of comment text because of its uneven distribution of features,a user's abnormal review detection model based on hierarchical multi-channel attention is proposed.we can grasp the local features of the review text by constructing multi-channel attention composed of emotional words,parts of speech and word positions.Then,CNN and Bi LSTM are merged to form a hybrid network,and different levels of user and product attention mechanisms are embedded,so that the global semantics of comment texts can be obtained at different levels.Finally,the model is compared with six comparison models on yelp dataset,and the experimental results show that the accuracy of the model is higher than that of the six comparison models,which proves that the model is effective for users' abnormal review detection task.2.Aiming at the problem of single feature input and different user behavior characteristics have different contribution to detection effect,a user's abnormal review detection model combining text features and behavior features is proposed.Firstly,this model obtains the semantics of review text through multi-channel hierarchical attention mechanism.Then,the user's behavior is expressed as a behavior feature vector,and the contribution degree of different behavior categories to the classification results is obtained by using the attention-weighted CNN network,so as to realize the abnormal review detection by using multi-dimensional features.Finally,this model is compared with eight comparison models on Yelp data set.The experimental results show that the accuracy of this model is higher than that of the eight comparison models,which shows the effectiveness of this model in detecting abnormal comments from users.