Research On Prediction Methods Of Click-Through Rate Using Multi-Scale Stacking Networks

Forecast for the click-through rate of advertisements is of great significance to plan for advertising,increase the income of the advertising platform and improve the user experience.How to get the rich and effective feature combination from the information of users,commodities,and their interaction in the data is the key to the prediction task.Generally,the dimension of advertising features is very high and there are three problems in that the features built are relatively simple,or the implicit semantic information of data is difficult to mine,or the combination of features is difficult to learn under sparse dataAiming at the above three problems,this paper designs a Multi-Scale-Stacking neural network structure,by adding the deep neural network and the factorizer parts to the model progressively we propose three click-through rate prediction models:MSSP-net,sDeepFM,and sDeepFM-2.The main research contents and innovation are as follows:(1)A novel structure named Multi-Scale-Stacking Pooling(MSSP)based on different receptive fields for multi-scale feature extraction is proposed,which mines high-order and low-order features in different local information from the directions of depth and width to ensure the diversity of extraction features.Comparing with the popular models LR,FM,and AFM,MS SP-net get greatly improved(2)To mine the implicit semantic information of data,the deep neural network is introduced into MSSP-net.We name the new model sDeepFM which combines the MSSP and DNN structure.Compared with MSSP-net,sDeepFM has achieved further improvement.(3)The second-order combined feature contains the largest amount of information for the prediction of the click rate of display ads.We carefully learn second-order combined features through a factorization machine,splice the learned features into sDeepFM to further improve the model and name the new model sDeepFM-2.The comparative experiment on two public advertising data sets shows that compared with three current excellent models PNN,xDeepFM,and DeepFM,sDeepFM-2 achieved better results on the indicators of Area Under the Curve(AUC)and Log-likelihood Loss(LogLoss).