Research On Fake Review Detection Method Based On Knowledge Graph

At present,e-commerce plays a very important role in daily life.And online reviews play a vital role in e-commerce applications,which could help people compare the quality of products,evaluate the service of stores,and serve as a basis for consumers to make purchasing decisions.However,driven by the interests,merchants began to hire water armies to impersonate ordinary customers to forge comments,trying to mislead consumers through fake comments.On the one hand,they praise their products,on the other hand,they maliciously slander their competitors.The existence of fake reviews poses a huge challenge to creating a fair and equitable online shopping environment.Therefore,how to effectively identify fake reviews becomes one of the network security issues that need to be solved urgently.Many workers focus on the detection of fake reviews,and the results of the research are beneficial to provide reasonable shopping decisions while guiding the benign competition of the business.Although researchers have made great progress in fake review detection,there are still many shortcomings.First of all,the results of researches did not explore the multidimensional features of the review text and the consistency between the review text and the scores.Secondly,the number of reviews and the implicit relationship between the scores and the time series were ignored.Finally,the influences about multi-mode network features identifying fake reviews were not considered.Therefore,the current algorithms about identifying fake reviews still have many problems.In view of the above problems in fake comment detection,this paper proposes and deeply studies a method based on knowledge atlas for fake review detection.The main work of this paper is as follows:(1)A fake review detection method based on scoring-text consistency was proposed.Firstly,we analyzed the emotional polarity of the comment text and consider the influence of emotional intensity and feature influence on the text polarity,further judge the consistency between the text polarity and the score.Secondly,we simplified the feature set by analyzing multi-source features,and then extracted five important features about fake review detection.Finally,we constructed a classifier for fake review detection by fusing multiple features.This method is an important prerequisite for calculating the authenticity of reviews in knowledge graph.(2)A method of fake review detection based on multi-dimensional time series was proposed.Firstly,on the two dimensions of score and numbers of comments,a time series curve was fitted by Bayesian algorithm.Secondly,we set sliding time window and use template matching algorithm to detect the fitting curve burst mode.Finally,the consistency between the score of burst time period and the number of comments was compared.This method is based on the characteristics of the number of reviews and the sudden rise or decline of the rating.It can not only detect fake reviews efficiently,but also detect the problems of the shop brushing effectively.(3)A method of fake review detection based on dynamic knowledge graph was proposed.Firstly,a multi-granularity bidirectional LSTM(ST-BLSTM)network model was constructed to extract four kinds of entities including reviewers,reviews,commodities and stores.Secondly,we defined the relationship measurement between entities and discussed the relationships among the four types of entities.Secondly,the influence of time factors on fake reviews was discussed,and we designed an iterative model which could add time features during the process of relationship extraction,and further constructed a dynamic graph network.Finally,four new indicators were defined to effectively measure the interaction among the four types of nodes,then a fake review detection classifier was constructed.In addition,for the sake of supporting the innovation of this paper,a large number of multi-feature comment data sets were collected using Octopus data processor.The data sets make up for the shortcomings of single-mode data and imperfect information.The data sets contain multiple features including user,commodity,review,store and comment time.The main steps of the data acquisition process are as follows: firstly,the rationality of the information sources was verified by the characteristics of the current fake reviews and literature search.Secondly,We standardized the raw data and perform multimodal features correlation analysis.Finally,the importance of multimodal features to fake reviews detection was verified.