The Design And Implementation Of CTR Prediction Model For Display Advertising

In recent years,with the rapid development of the Internet,information dissemination has become the main mode of Internet communication.Internet advertising,as a new way of disseminating information,has had a profound impact on the development of the Internet and has become one of the major ways for most Internet companies to make profits.The expected cost per mille(eCPM)of thousands of impressions in the internet company's advertising business is one of the most important metrics for measuring revenue.It is expressed as the product of clickthrough rate(CTR)and click value,and the click value is already at the beginning of advertising.OK,so click-through rate becomes a particularly important parameter.For advertisers,the advertising and promotion of advertising products are improved;for advertisers,more revenue can be obtained;for the majority of users,the user experience can be improved and advertising information can be better targeted..This thesis focuses on the CTR prediction model based on display ads.Different from traditional search advertising,the click data of display advertisements is more sparse,and the long-tail effect is also more obvious.In the prediction of click-through rate,two different forms of discrete features and continuous features need to be considered at the same time,resulting in an accurate pre-click rate.Estimates are more difficult.In order to more efficiently and accurately predict the click-through rate of advertisements for these issues,the paper proposes the idea of model combination,that is,to establish different types of decision trees in the feature extraction phase so that the number of exposures for long-tail advertisements and new advertisements can be reduced.Get sufficient training;At the same time,for the two forms of features,the continuous features are discretized through the decision tree,then combined with the discrete features and encoded into the FM model to get high-order features,and finally input into the LR model for training.Based on the design and implementation of the model,the thesis implements a prototype system and uses the Logistic Regression,Factorization Machine,and Gradient Boost Decision Tree models commonly used in the industry.The functional and performance experiments are evaluated.The function is mainly based on the contrast between the AUC(Area Under Curve)and online CTR values.The performance is mainly based on model iteration time,overall training time,CPU occupancy ratio,etc.Performance indicators to judge.The experimental results show that the proposed combination model can better meet the business requirements and is more accurate than the traditional prediction model.