Research On Advertising Click-Through Rate Estimation Technology Based On Feature Learning

The issue of click through rate estimation in computational advertising has been paid close attention in the academic and industry areas.It has been widely studied in information retrieval,machine learning,query recommendation etc.Click through rate estimation in sponsored search is the task of predicting probabilities that user click an ad given<query,ad>and context information.Studies about this issue in the academics mainly have two approaches.First,statistical learning model,which by designing the feature extraction scheme is a key part of program,such as extracting relevant features between advertisement in the same resulted page or construct combined features and so on,as characterized by obtaining highly correlated with CTR model,thus improving the accuracy of the model estimation.Second,user behavior modeling based on Probabilistic Graphical Models,through hypothesis testing,take advantage of Bayesian network to portraying user's browsing scenes,and then estimate the probability that the user clicks an ad.Human feature engineering construct combined features,it exists low efficiency,scalability and performance problems and other issues.While Bayesian model characterizes user browsing behavior,it exists use of information is insufficient,and don't take into account the sparseness of advertising data and highly nonlinear association between single features.Against these issues,this paper considers the characteristics of advertising data and proposes advertising data-oriented sparse feature learning from the perspective of learning characteristics to estimate CTR.This method combines the advantages of tensor dimensionality reduction and feature learning to solve high-dimensional sparse feature problem of advertising data.Firstly,there is a relationship between internal objects of the same type,dimension reduction using clustering makes an initial aggregation of data;as for association between different types of objects,dimensionality reduction using tensor decomposition,while protecting the data associated with the original structure of ad clicks,lower characteristic dimension.Secondly,a stacked auto encoder algorithms in deep learning fields has been studied to mining high-order relationship between advertising data features,and obtaining new abstract features,which has more representation capability to ad data and help to improve the prediction accuracy of the CTR.Thirdly,the new features will be acquired as the input click prediction model and use the L-BFGS algorithm to learn parameters.Finally,in the experimental part,this paper compare with the existing methods to verify the estimated effects.Experimental results show that the method this paper proposed can promotes the estimation accuracy dramatically compared to existing method.