Research On CTR Prediction Based On Deep Learning

With the rapid development of the Internet,Cloud Computing,and Internet of Things,the data of the network is also growing rapidly.Actually,the information society has slowly entered the "big data" era.Network advertising system Based on the big data utilizes the machine learning method to fully exploit the massive user behavior and push the appropriate advertisement to user at the right time.Accurate prediction of CTR is to develop a scientific e-commerce marketing decision-making key,affecting the user’s network experience,and is directly related to the operating costs of Internet companies.Therefore,the prediction of CTR has a high commercial value and research value.In the face of the high precision and high efficiency requirements of the online advertising system,this paper carries out the feature selection,feature learning,classification prediction and the application technology study from shallow learning and deep learning.Taking the real data set of online advertising as the experimental object,shallow learning model and deep learning model are constructed respectively in this paper.In order to fully validate the deep learning,this study confirms the great potential of deep learning through a multi-perspective comprehensive comparison experiment.Taking into account,the specific research work mainly includes the following five aspects:(1)Do the research on data processing and feature engineering technology.Explore the impact of category imbalanced-data on prediction model from the real dataset and the resampling technology of imbalanced-data.(2)Aiming at the highly nonlinear features of data,a comparative study of shallow learning,deep learning theory and application technology is carried out.In order to overcome the problem of limited ability of complex problems,this paper constructs deep learning model.The experiment proves that the prediction effect isabout 21% higher than that of shallow learning.Deep Learning has a strong advantage.(3)In order to eliminate the influence of class imbalance on the prediction model,an improved model of Deep Neural Network(DNN)is proposed,which is the SDNN(Deep Neural Network based on Sampling,SDNN).Based on the parallel computing of GPU,the training time of SDNN prediction model is shortened by about 73.28%without significantly affecting the forecasting effect by constructing model and implementation algorithm,which greatly improves the efficiency of DNN.SDNN has been proved to be a more efficient forecasting method of big data for the high accuracy and timeliness of the system.(4)To study the influence mechanism of Sigmoid activation function and Relu activation function on DNN prediction model.By constructing the DNN and SDNN models and algorithms respectively,it is proved that the Relu activation function is more suitable for the deeper network model than the Sigmoid activation function.DNN and SDNN based on Relu activation function are more suitable for the modeling of complex problems.(5)In order to avoid the limitations of single SDNN training and enhance the generalization ability of the model,the study of key parameter dropout sensitivity analysis is carried out.