Research On Spam Detection In Reviews Based On Feature Fusion

Due to the rapid development of Internet technology and the change of global consumption pattern,it has become normal for people to comment on the products or services they buy.As reviews increasingly influence people's buying decisions,voting or product's design,it has become a common practice for spammers to write negative reviews on competitors' products or give favorable reviews on their own.Spam reviews may distort opinions,create public opinion,and reduce the credibility of the review system.However,it is unrealistic to identify spam reviews only by manual observation,so it starves for a method,which can make the computer automatically conducts statistical analysis on the massive review data,and then judge whether the reviews are spam or not.Thus,review spam detection came into being.Currently,the existing researches on review spam detection always cannot fully consider the temporal features of reviews in multiple entities,cannot effectively integrate the independent text and behavior features of different entities,and provide poor detection performance under the cold start condition.To address above problems,this thesis takes advantage of deep neural network to carry out in-depth research on review spam detection.The main research results are as follows:(1)A review spam detection method using LSTM-based multi-entity temporal features is proposed.Current works on spam detection in product reviews tend to ignore the temporal relevance among reviews in the user or product entity,resulting in poor detection performance.To address this issue,the present paper proposes a spam detection method that jointly learns comprehensive temporal features from both behavioral and text features in user and product entities.Firstly,the behavioral features of a single comment are extracted according to the expert knowledge,and employ a convolutional neural network to learn the text features of this review.Then,combine the behavioral features with the text features of each review and train a Long-Short-Term Memory(LSTM)model to learn the temporal features of each review in the user and product entities.Finally,train a classifier using all the learned temporal features in order to predict whether a particular review is spam.Experimental results demonstrate that the proposed method can effectively extract the temporal features from historical activities,and can further jointly analyze the activity trajectories from multiple entities.Thus,the proposed method significantly improves the review spam detection accuracy.(2)A review spam detection method using GCN-based multi-entity temporal features is proposed for the cold start problem.The existing review spam detection methods cannot make full use of the association information between users,which leads to the lack of features in the cold start condition.A new review spam detection method focusing on the cold start problem is proposed.Firstly,the behavior features of each entity and the text features of all reviews are extracted,and the user entity and the product entity with features are employed as nodes to construct the heterogeneous graph,these nodes are correlated according to the user's review of the product.Then,the Graph Convolutional Networks(GCN)is employed to learn the shared behavior features among the nodes in the constructed heterogeneous graph,which is used to enrich the behavior features of cold-start users.Finally,a classifier is constructed by combining the shared features,text features of review and behavior features of all entities to realize the detection of spam reviews under the cold start condition.The experimental results show that the shared features learned by GCN can effectively promote the detection performance in the absence of user behavior information.Compared with the existing methods,the proposed method improves the overall detection performance of spam reviews under in the cold start condition.