Research On Optimization Method Of Click-through Rate Estimation Based On Wide & Deep Model

Click-through rate estimation,as a key research issue in the Internet industry's traffic distribution link,is significant to the business of content recommendation,online advertising and other fields.Recently,deep learning technology has powerful information mining capabilities in the field of big data,which makes it widely used and excellent in the problem of matching lots of users with a large amount of content.This paper combs the research results of various CTR estimation by means of deep learning in recent years.Firstly,it analyzes the characteristics of the joint training framework of the Wide & Deep model and focuses on the characteristics of its balanceable model memory and generalization.Secondly,it summarizes the advantages and disadvantages of classic algorithms in user historical preference information mining and finds that they do not consider the target items and the time sequence information of user behaviors at the same time in the use of user historical behavior data.Matching user interests is very important.Finally,this paper designs and implements a sub-network structure based on the multi-head self-attention mechanism and shallow feedforward neural network that captures the user's historical preferences,and obtains a new CTR estimation algorithm called BSN-FM.The main content is as follows:Firstly,the key to analyzing the CTR estimation problem is to predict the degree of matching between user interests and the items to be estimated.This paper sorts out the advantages and disadvantages of the CTR estimation algorithm based on the Wide &Deep model.It also summarizes the application of the Multi-head Self-attention mechanism in the field of NLP and CTR estimation.Secondly,based on the research and improvement of the application of the ulti-head Self-attention mechanism in the Seq2 Seq scenario,a network structure is designed to mine the relationship between the advertisements to be estimated and the user's historical preferences.A new algorithm called BSN-FM is established,which is based on Wide &Deep model.Thirdly,in view of the typical attributes and common content in the click-through rate data set,the relevant content of Bayesian statistical theory is introduced,and a complete and generalizable method of constructing the characteristics of the clickthrough rate data set is designed and applied to the selected experimental data set.Finally,a controlled experiment for the selected experimental data set is constructed.And the effectiveness of the BSN-FM model is verified through the experiment.The AUC of the BSN-FM model multi-round training reaches 0.8328.The AUC of the singleround training AUC reaches 0.8085.Compared with other classic CTR prediction models,the AUC is relatively increased by 2.41%.