Variable Selection Method In Logistic Model And Its Application In The Search Engine Advertising Conversion

Logistic model is an important method to study classification variables.The data in the big data era is high-dimensional and sparse,so it is necessary to study variable selection based on Logistic model using the high-dimensional data.In this paper,the variable selection method in the Logistic model,included MCP(Minimax Concave Penalty)and SLS(Sparse Laplacian Shrinkage),and its application are analyzed.The specific research contents we have carried on and the main results are as follow.The paper constructs the probabilistic relation among the factors that affect the conversion of search engine advertisement(SEA)using the probabilistic factor graph.Then the Logistic model based on multi-layer priori method is established to predict the probability of SEA conversion.The model is proved to be reliable by the empirical analysis.The analysis also show that the advertising quality has a greater impact on the SEA conversion rate among the influencing factors,and advertisers can improve landing page quality and user experience by optimizing website design to obtain more advertising conversion.It is proved that the MCP estimation of parameters in the Logistic model has Oracle property under regularization condition.The MCP method is applied to the SEA conversion rate prediction and the variables which have significant influence on the conversion are screened out.Analysis shows that the advertising position ranking,keyword information and advertisement landing page quality all have a significant impact on the conversion rate.Advertisers can improve conversion rate by selecting appropriate keywords,improving website access speed,etc.For the case where there is a correlation between variables,SLS method based on Logistic model is proposed in this paper.The main principle of this method is to represent the network structure between variables by constructing the adjacency matrix,and the network structure penalty item is introduced on the MCP penalty.In addition,this paper also introduces two simple methods of constructing adjacent matrices based on similarity measure and dissimilarity measure.Finally,the numerical simulation verifies that compared with MCP,the SLS method has a better recognition accuracy of effective variables and prediction when the variables are related.