Research On Click-through Rate Prediction In Sponsored Search Advertising Based On User Feature

With the rapid development of science and technology,Internet ads has become a trend in the global scope.As major revenue source of search engine company,sponsored search advertising has become the hot research of Internet ads field in recent years,and click-through rate prediction in sponsored search advertising is key issue in this research direction.lots of logs are generated when users interact with the search engine,this paper use these message to build feature system.Combined with advantages of existed prediction model,putting forward fusion technology based on rank factor then applying to CTR prediction.According to poor learning ability of shallow model,putting forward CTR prediction based on deep neural network.This paper focuses on feature related to CTR,build feature system with high influence factor,This paper focuses on feature related to CTR,and build feature system with high influence factor,and use LDA to extract topic from user query,define the method to extract similar feature and statistical feature.Then,this paper build CTR prediction model based on online Bayesian probability and logistic regression respectively.According to their different performance in different data sets,analyses their advantages and disadvantages.Combined model the advantage of dealing with miss information in online Bayesian probability and feature sensibility of logistic regression model,we put forward fusion technology based on ads rank.Finally,according to the poor ability of shallow model,this paper pick sigmoid function as activation based on the artificial neural network and back-propagation neural network,put forward building method of multi-level deep neural network and control error range under threshold,then improve prediction accuracy.This paper pick the KDD Cup 2012 Track 2 data sets,and analyses this data sets.The user feature based on LDA has more influence on CTR prediction,and fusion model has more excellent performance than single shallow model.Deep neural network has more advantages than logistic regression model.