Advertising Click Through Rate Prediction Algorithm Based On Deep Learning

Internet advertising has become the main source of revenue for the Internet Co,and the prediction of the rate of click on ads is the most important job,and the accuracy of increasing the click rate can directly bring income to the company.Advertising log data and images or audio continuous features are different,advertising data are discrete,high dimensional,and their dependence is very small.At present,the mainstream method,such as Baidu and Google,is the linear model adding a large number of artificial features,and this method is more and more difficult to follow,because a large number of manual features need to spend a lot of manpower,and the benefit is not improved.The reason is that these characteristic linear models can not learn the nonlinear relationship between the characteristics.By using the conventional deep learning model,the deep learning network can not converge because of the large amount of computation caused by the large input space.In order to solve the above problems,this paper proposes a new deep learning model that can deal with a large number of sparse data,and can effectively learn the nonlinear characteristics of the data.Large-scale experiments and real data show that the model proposed in this paper can be significant.Improve accuracy and recall.The research work of this article focuses on four aspects:First: Improved data embedding algorithm,aiming at the characteristics that the real data ID class data is particularly large,study and improve the embedding algorithm so that the original data will not suffer loss after embedding the algorithm,and reduce the dimension of the input vector so that deep learning can be processed.Secondly,Improved depth learning model named BFNN(Boosting tree factorization-machine supported Neural Networks)is proposed.Based on the neural network of lifting tree and factorization machine,the deep learning model proposed in this paper is due to the fact that the real data has high-dimensional nonlinearity.Can effectively learn the nonlinear characteristics of the data.Third: The difference between the deep learning model and the shallow model in the CTR estimation problem.Fourth: Verify the effectiveness of BFNN in improving the number of network layers,network structure,regularization and activation functions.