Research On Fake Review Detection Method Based On Deep Transfer Learning

Online reviews play an important role in e-commerce.Users need to check relevant reviews before purchasing products,and these online reviews also directly affect the reputation and profitability of merchants.With the rapid development of e-commerce,fake online reviews are increasing day by day,which seriously interferes with consumers’ shopping decisions and pollutes a fair e-commerce environment.Although there are many fake review detection methods the detection results are not satisfactory.First,the feature representation of fake reviews is not rich enough,the multi-modal features of reviews are not fully utilized,and the interpretability of detection results is lacking;Second,the positive and negative samples of the review dataset are not balanced,which seriously affects the detection results of fake reviews;Finally,the current popular self-supervised deep learning detection methods do not make full use of known labels and cannot effectively distinguish fake reviews with similar semantics but different categories.In response to the above problems,this thesis studies the characteristics of fake review data and existing fake review detection methods,and further proposes a series of methods based on deep transfer learning.The main contributions of this thesis are as follows:(1)An interpretable fake review detection method based on knowledge integration(EKI-SM)is proposed to solve the problems of insufficient feature representation of fake reviews and lack of interpretability of detection results.First,domain knowledge is integrated to guide fake review detection while providing interpretability for detection results;second,multi-modal features are fused,using 1D convolution,Long Short-Term Memory Network(LSTM),skip layer,and other methods to learn high-dimensional feature representation of fake review texts,which effectively improves the detection performance of fake reviews;Thirdly,inspired by interpretable deep learning,EKI-SM can discover important words in review texts,which provides explanations for fake review detection results.(2)A fake review detection method based on data augmentation and model transfer(DA-BERT)is proposed to solve the problems of unbalanced positive and negative samples in the review datasets and low accuracy of detection results.First,we balance the positive and negative samples in the datasets with two data augmentation methods,EDA(Easiest Data Augmentation)and back-translation;Secondly,we transfer BERT(Bidirectional Encoder Representation from Transformers)pre-training model to learn the reviews.Furthermore,compared with the latest fake review detection methods,our method in this thesis can effectively balance the positive and negative samples of the datasets and improve the accuracy of prediction.(3)A fake review detection method based on fine-tuned transfer model(CL-BERT)is proposed to solve the problem that self-supervised deep learning detection methods do not fully utilize known labels and cannot effectively distinguish fake reviews with similar semantics but different categories.First,a BERT pre-training model is used to obtain the vector representation of reviews;Second,a supervised contrastive learning fine-tuning model is used to enhance the vector representation of fake reviews with similar semantics but different classes;Third,compared with traditional deep neural network models,our proposed method has a strong classification ability.We conduct extensive experiments on multiple public datasets and use multiple evaluation metrics for evaluation.The results show that the series of fake comment detection methods proposed in this thesis has better performance and provide the interpretability of the detection results as well.The detection results reach or exceed the current mainstream fake review detection methods.