Research On Fake Review Detection Method Based On Sentiment Analysis And Hybrid Sampling

Online reviews of e-commerce products or services have a significant impact on consumers’ purchase decisions.At present,out of a profiteering attitude,some merchants fabricate a large number of fake reviews.It not only seriously hampers consumers’ normal judgment,but also further undermines the fairness of the online business environment.Tackling the proliferation of fake reviews becomes an urgent issue for both academia and industry.On the one hand,it can help e-commerce platforms to identify merchants’ violation behaviors and therefore take effective actions for internal governance timely.On the other hand,it can promote the platform to improve the rules or mechanisms,so as to guide businesses to avoid the profiteering attitude driven by interests.Due to the extensive participation of consumers,online reviews are characterized by large amount of data,complex features,and unbalanced categories.Therefore,how to identify fake reviews from large-scale online review data has important research value.For the detection of fake reviews,scholars have conducted extensive and in-depth research,and have proposed various efficient and accurate models or methods.However,there are still a lot of problems worth exploring,such as the extraction of emotion features and the solution of data imbalance.On the one hand,studies have proved that the fusion of emotion features can effectively improve the performance of recognition algorithms,but most of the existing methods use coarse-grained emotional polarity features(sentence level or review level),and fine-grained emotion features(such as aspect-based sentiment features based on product attributes)have not been fully utilized.On the other hand,a large number of studies have proposed many methods to deal with the class imbalance problem,but the class overlap problem has not been paid enough attention.Both the class imbalance problem and the class overlap problem will affect the classification accuracy of the model,and the coexistence of the two types of problems brings greater challenges to the model design.In order to solve the problems described above,this dissertation conducts a study of the method of fake review identification by adopting fine-grained sentiment feature extraction technology and new hybrid sampling technology.The main innovations of this dissertation are as follows:First of all,in order to solve the problem that fine-grained emotion features have not been fully explored and utilized in fake review detection,this dissertation proposes a heterogeneous ensemble method for fake detection based on aspect-based sentiment features.Aspect-based sentiment feature is a fine-grained emotion feature based on product attributes.It can more specifically and accurately depict the differences of sentiment between real reviews and fake reviews,so as to help identifying fake reviews effectively.Firstly,fine-grained sentiment features on seven dimensions are obtained by lexical analysis and lexicon-based sentiment analysis.These features are integrated with review text features and user features.Then,the heterogeneous ensemble learning method based on soft voting strategy is used for classification prediction,and the effectiveness of the proposed method is verified by a large number of comparative experiments.Secondly,taking consideration of data class imbalance,this study further considers the problem of class overlap,and proposes a fake review detection method based on hybrid sampling of class overlap degree.This method first uses a K-nearest neighbor algorithm to define the class overlap degree,and conducts down-sampling for majority class samples based on class overlap degree ranking,and then uses the SMOTE method to up-sample minority class samples,so as to obtain the class-balanced training data set.This hybrid sampling method can minimize the influence of the distribution changes of data caused by sampling on the model effect.Then,the samples are classified by the ensemble learning method which can adjust the sampling parameters adaptively.Finally,experiments show that this method can significantly improve the detection effect of fake reviews.Thirdly,a fake review detection method based on sentiment analysis and hybrid sampling is proposed,by fusing the heterogeneous ensemble method of fake review detection based on aspect-based sentiment features and the ensemble method of fake review detection based on hybrid sampling of class overlap degree.Experimental results show that this method has more advantages than any single method in the detection of fake reviews,and can comprehensively improve the detection performance.Therefore,this method can simultaneously solve the problems of the under-utilization of fine-grained sentiment features and the coexistence of class imbalance and class overlap.In conclusion,aiming at extracting the hidden fine-grained emotion features and solving the problem of the coexistence of class imbalance and class overlap in online review data,a fake review detection method based on the fusion of aspect-based sentiment features and a fake review detection method based on the hybrid sampling of class overlap degree are proposed,respectively.The above two methods are combined to form a fake review detection method based on sentiment analysis and hybrid sampling.This method can help e-commerce platforms detect fake reviews more effectively from large-scale online reviews,and has potential value in managerial practice.