Boosting-based Method For Advertising Conversion Rate Prediction

Internet advertising conversion rate is an important quantitative indicator for search engine service providers and advertisers,the realization of the prediction of Internet advertising conversion rate under the big data platform has strong theoretical research value and practical application value.The conversion of Internet advertising is a small probability event under massive data,by estimating the conversion rate of advertisements,it can not only improve the effectiveness of advertisements,but also turn advertisements from useless harassment information into valuable information,thus bringing great convenience to users' work and life.Therefore,it's important to increase advertising conversion rate through forecasting.This article aims to apply efficient algorithms and models to the prediction of advertising conversion rates by modeling user portraits,advertising messages,and contextual information.The specific research content is as follows:1)For the ad conversion log and external feature files,data mining of the conversion log is achieved through the mining of features of the calculated advertising data.Data mining is performed on the main role users of clicks and conversions of ads,and an extraction method based on user's click features is proposed.Content includes basic feature extraction,user statistical feature extraction,smooth conversion rate extraction,and other extraction methods.A new intermediate dataset is generated for each feature extraction.Feature engineering is used to extract,synthesize and transform features etc.,and gradually improve the training set and test set required for prediction.2)Aiming at the characteristics of high dimensionality of advertising conversion data,a Gradient Boosting Decision Tree(GBDT)forecasting method for advertising conversion rate is implemented.GBDT is a typical boosting tree algorithm based on Boosting idea.It has innate advantages for advertising a large number of high-latitude feature combinations.Since GBDT is easy to overfit,a feature combination method is proposed: Combining Logistic Regression(LR),which excels at dealing with a large number of discrete features,with GBDT models improves the fitting of model to improve accuracy of the prediction.3)Due to low efficiency and weak scalability for GBDT,Two GBDT-based gradient boosting models for e Xtreme Gradient Boosting(XGBoost)and Light Gradient Boosting Machine(Light GBM)are used to estimate the advertising conversion rate.In order to further improve the accuracy of prediction,a prediction method for advertising conversion rate based on dual model fusion is proposed.The two single models were trained using the Stacking and Blending model fusionmethods respectively,and better prediction results than the single model were obtained.