Research On The Methods Of Fake Ratings Detection And Reputation Evaluation In E-Commerce

In recent years,with the rapid development of global e-commerce industry and the continuous expansion of industrial scale,it has occupied an increasingly important position in the national economy.However,driven by the business interests,many dishonest serllers take advantage of the information asymmetry of e-commerce platform and the weakness in the lagged development of the reputation system,use various means to destroy the reputation system,mislead consumers and e-commerce platform,which has a negative impact on the healthy development of e-commerce industry.In view of the fragile reputation system of e-commerce platform and the problem of repeated fake ratings,this dissertation starts with the rating behavior of the buyer(reviewer),and studies the buyer classification method,the reputation evaluation of c-commerce platform and the detection of the buyer's fake ratings.The major research works are as follows:1.Based on the impression theory,this dissertation models the rating behavior of buyers,defines the lenient and the strict buyers,and designs the nearest neighbor search method based on the behavioral characteristics of buyers.This method improves the nearest neighbor search algorithm,because it does not spend extra search time on the cluster of remaining buyers,so it improves the speed of classification.The experimental results show that the classification method is better than the traditional classification methods.2.An unsupervised sellers' reputation evaluation method IBS(impression-based strategy)based on buyer's behavior is proposed.First of all,according to the behavior characteristics of the strict and the lenient buyers,the evaluation rules of sellers' honesty is put forward.Secondly,the rules of sellers' attribute evaluation are used to pre-classify some honest and dishonest sellers.Based on those sellers,the buyers are divided into three categories:honest,dishonest and uncertain.Finally,to calculate the seller's reputation,the reputation score of honest and uncertain buyers is weighted and aggregated.Experiments on simulated datasets and Yelp datasets show that IBS not only can accurately estimate the seller's reputation,but also can defend multifarious common and unknown reputation attacks.Even in the extreme environment with a high proportion of dishonest buyers,IBS can also works effectively.3.Based on deep learning theory,a semi-supervised fake ratings detection algorithm is proposed.The algorithm uses a Markov Decision Process to model the buyer's ratings sequence,and designs a deep Q network based on the buyer's rating characteristics to learn the buyer's behavior.In order to perceive the change of environment more quickly,the seller reputation evaluation method of IBS is introduced into the deep Q network.Experiments based on real data show that the improved deep Q network,that integrates the seller reputation evaluation method,can be used as a filter to detect the authenticity of other ratings in the same platform after about 20000 samples of training and learning.4.This dissertation designs and implements an e-commerce reputation simulation platform based on multi-agent theory and technologies.The platform takes into account various issues affecting the market,such as reputation and price factors,trading and selection strategies,as well as the behavior patterns of all participants in the market.Both distributed and centralized reputation models can be loaded,which facilitates the comparison of reputation models.