Research On The Click-through Rate Estimation Of Display Ads Based On The Multi-classifier Fusion Model

The mainstream trend of today's Internet advertising is “precise”.The intelligent marketing platform has accumulated a large amount of advertising data and user data.How to effectively use this data to predict the user's advertising click probability is the key issue of the application of big data in precision marketing.This paper is mainly based on feature engineering and model construction.It builds a display ad click rate prediction model,which has important guiding significance for advertisers to achieve accurate delivery,maximize benefits and improve users' online experience.Feature engineering.First,a series of visual exploratory analysis is performed on the advertising data,specifically through pie charts,stacked charts,boxplots,and column charts,etc.,to directly observe the structure and characteristics of the data set,and initially verify that the click rate distribution of different categories of features is significant difference.Next,data cleaning,feature reduction,and feature transformation are performed on different types of features to effectively reduce data noise.Finally,from the three aspects of feature selection,feature extraction,and feature construction,the multi-source features such as creative information,user information,context information,and media information are fused and extracted to obtain a large amount of effective feature set data for subsequent machine learning algorithms,which can be able to provide a solid foundation for better performance.Model building.First,using the original data set and the data set processed through feature engineering,three single algorithm models,Logistic regression,XGBoost model,and LightGBM model,were trained in turn.By comparing LogLoss and AUC,the prediction effect of the LightGBM model based on feature engineering was obtained the best.Then,based on the Stacking integration idea,a multi-classifier fusion model that predicts the click-through rate of display ads is proposed.Through real advertising desensitization data,it is obtained that both LogLoss and AUC methods are significantly better than the above single algorithm model evaluation.The validity of the model has certain extension significance for the application research of display ad click rate.