Research And Implementation Of Mobile Advertising Click Through Rate Prediction

The Click Through Rate(CTR)prediction is one of the core issue of online advertising,which directly affects the income of the search engines,social platforms and other intermediate platforms and advertisers.The traditional CTR prediction model of advertisement based on PC terminal has been more mature,along with the arrival of internet time,the mobile terminal ad CTR prediction model still follows the traditional PC model.The mobile terminal is different from PC terminal on usage environment,display forms of advertising and other aspects,the mobile terminal ads and PC terminal ads also differ in data characteristics and model applications.Therefore,this thesis studies the CTR prediction of mobile advertising,the main work is as follows:1.Based on the analysis of the differences between mobile ads with PC terminal ads and the characteristics of mobile ads,the representative features are selected and the mobile user preference characteristics based on time are extracted.2.The CTR prediction model of Field-aware Factorization Machine Based on Gradient Boosting Decision Tree(FFM++)is researched and implemented.Firstly,FFM++ model finds a variety of distinguished features and feature combinations by using Gradient Boosting Decision Tree,because Field-aware factorization machine model has advantages in dealing with sparse data and multi value classification features,it is used to forecast the mobile advertising click rate.The results on Avazu datasets and Criteo datasets show that the proposed FFM++ model performs well compared to the model using gradient boosting decision trees and other basic models.3.A CTR prediction model based on reinforcement learning backpropagation neural networks is implemented.The neural network has good ability in nonlinear fitting,but it is not good at dealing with the multi-value classification feature after one-hot encoding.It is a process of gradually approaching the optimal strategy by learning experience through actual system.So a RBP model based on improved Field-aware factorization machine has been introduced in this thesis.Firstly,the model map the input by using learned neural network,then a reasonable evaluation to output of network is given by the modified Q-learning algorithm which altered through neural network,finally,the evaluation result will be submitted to the network in order to adjust the weights of neural network.The experimental results on Avazu dataset and Criteo dataset show that the effect of RBP prediction model is not only better than that using neural network model alone but also better than that of FFM++ model.