Research And Implementation Of Fake Review Detection Based On Deep Learning

Authentic and credible reviews in e-commerce platforms can help consumers make correct decisions.However,driven by some reasons,reviews in e-commerce platforms contain lots of fake reviews which will mislead users' consumption decisions,and have a negative impact on businesses and e-commerce platforms.On account of that,the detection and regulation of fake reviews is of great significance for the supervision of platform operation and the purification of network environment.The fake review detection method proposed in this dissertation is based on deep learning.More specifically,in order to ensure the accuracy and efficiency of this system,two fake review detection methods are proposed from the perspective of review similarity and review sentiment.And further,to fully exploit their advantages,a new method that considers multiple clues is devised,and a system which can automatically collect reviews and then detect the fake ones is designed and implemented.The main contents of this dissertation are listed as follow.(1)A fake review detection method based on graph convolution network is proposed.Fake reviews often have highly similar content,while the existing methods of fake review detection based on deep learning usually extract features for each review separately without fully considering the similarity between reviews.In this dissertation,the semantic similarity of words in reviews is used to measure the similarity between reviews indirectly,and based on that,fake reviews are detected with a graph convolution network.Firstly,a review text graph is constructed based on lexical semantic similarity,then the fake review detection problem is transformed into a node classification problem.Secondly,the graph convolution network is used to aggregate the neighborhood information of adjacent nodes,extract the similarity relationship between reviews,and obtain the feature vectors containing the similarity between reviews for fake review detection.Experimental results on general fake review detection dataset reveals that,compared with CNN,LSTM and Text-GCN,the accuracy is improved by 7%,4.8% and 1.3% respectively.(2)A fake review detection method based on sentiment feature is proposed.In order to achieve the effect of propaganda,fake reviews usually convey strong emotions.Yet most of the existing detection methods based on sentiment features simply count the sentiment attribute words,ignoring the contribution of their intensity diversity.Therefore,a new detection method based on sentiment features is proposed.Firstly,sentiment dictionary is constructed based on the review text,and then the form divergence of sentiment words and the dependence on adverbs are comprehensively considered to describe the sentiment expressed in reviews.Secondly,fusing these sentiment features and text content features,and further using Transformer model to realize fake review detection.Experiment on Amazon dataset reveals that,compared with LSTM model,this method improves the detection efficiency and accuracy by0.59%.(3)Given that a fake review may have both of the characteristics mentioned above,the two fake review detection methods are fused into a new one that considers multiple clues,and based on this method,an online fake review detection system is designed and implemented.The system is divided into two modules: server and client.The server mainly realizes the core functions of review data collection,model training and reviews online detection;the client mainly realizes the management of review detection task and display review detection results.On Amazon dataset,the detection accuracy is improved by 3.24% and 1.14% respectively compared with the two methods mentioned above.The experimental results show that the fusion of the similarity and sentiment features of reviews is effective.And moreover,on Dianping platform,it is verified that the system can realize the automatic collection of reviews and online detection.