Research On Mobile Advertising Click Fraud Detection Based On Ensemble Learning And Deep Sequence Model

With the popularity of smartphones,advertising industry relies on the accuracy,timeliness and various forms of mobile Internet and developed rapidly.However,the cost per click(Cost per click,CPC)payment model was maliciously exploited by black markets,advertiser budgets were consumed,which severely affected the development of the mobile Internet advertising industry.In addition,there are crowdsourcing and incentive click fraud in ads click fraud scenarios,that is,by distributing click tasks with rewards to real users or motivating users to click ads that they are not interested in,therefore it is harder to detect fraud because they are real clicks.This paper aims at the task of detecting click fraud in advertising agencies on mobile ads platforms,and uses machine learning methods based on ensemble learning and deep sequence model to solve this task in different scenarios.The differences in users' behavior patterns were excavated and detection accuracy was improved.The main work of this paper is as follows:(1)Based on the users' click sequences,the Fb2vec(Fraud behavior to vector)algorithm is proposed.Item2vec is used to mine the publishers'embedding vectors representing users' preferences in users' click sequences,and two strategies are proposed in its loss function according to differences in users' click behavior patterns,namely "users' behavior pattern capture strategy" and "single click optimization strategy".The vectors visualization and multiple sets of control experiments shows the effectiveness of the algorithm,the accuracy is significantly improved by 0.08%?1.19%in fraud detection.(2)A click fraud detection method based on ensemble learning is designed,and a synthetic oversampling algorithm based on weak classification is proposed.Through pretraining,hard-to-classify samples are obtained and then synthesized for sampling,which solve the problem that SMOTE(synthetic minority oversampling technology)algorithm only samples minority and thus not being generalized well.The experimental results show that the accuracy is improved by 0.21%in ensemble learning detection based on the synthetic minority oversampling method.(3)For crowdsourcing and incentive ad click fraud scenarios,this paper uses the click sequences as the basis.First,it extracts Fb2vec features and four kinds of sequences were extracted,namely click traffic sequences,single click(the user which only clicked once)traffic sequences,two clicks interval and three clicks interval sequences.Then,a click fraud detection model based on LSTM Attention is designed,and based on this,a wide and deep sequence model(Wide&Deep sequence model,WDSM)is proposed.The control experiment results show that,as the time granularity becomes finer,the accuracy of the model becomes better.The detection results at different time intervals of WSDM improve by 0.73%?7.59%compared with the ensemble learning model.