The Study On Online Ad CVR Predicting Based On Ensemble Leaming

With the development of the mobile Internet,research in the field of online advertising has also taken more new directions.In the past two years,with the further development of user tracking technology,there has been a new way of charging for the number of conversions of advertisements.How to match the highest conversion rate for users has become a hot research direction in recent years.Ad conversion rate estimates have several difficulties:high conversion data dimensions,highly sparse conversion data,and non-random error tags in the sample,which makes the current ad conversion rate predictions still unsatisfactory.Based on the above background,this paper applies the integrated learning method to the estimation of online advertising conversion rate,and explores the conversion rate estimation and personalized recommendation under the multi-dimensional feature model.The research mainly includes two aspects:(1)feature engineering.This paper analyzes the characteristics of online advertising,establishes a multi-dimensional feature model of {user,advertisement,situation},and adopts three correction methods:Bayesian smoothing,extreme hot item penalty and natural attribute feature probability.The basic multi-dimensional feature model is modified.(2)Model algorithm.Based on integrated learning,a more complex three-layer ensemble model is designed.The model includes three sub-ensemble models:GBDT's Stacking ensemble model.FFM's Stacking ensemble model,and XGBoost and FM serial ensemble model.In this paper,the evaluation results of the two evaluation indexes of logloss and AUC are used to evaluate the model.Compared with the single model,the feature correction method and the integrated model proposed in this paper have a significant effect on the conversion rate prediction.Compared with the traditional method,the proposed method makes full use of the advantages of the tree model and the generalized linear model by integrating multiple models,which reduces the over-fitting phenomenon of the model and improves the generalization ability and the accuracy of prediction,an effective way to predict sparse data.Through the above research,this paper enriches the theoretical basis of online advertising personalized recommendation field,and also provides a more general method for mobile online advertising conversion rate estimation.