Research On Advertising CTR Prediction Based On Improved And Optimized Logistic Regression

With the rapid development of Internet technology and the arrival of big data era,the mode of internet+marketing is more and more favored by Internet companies,so Computational Advertising comes into being in a new form of advertising.Advertising click-through rate is the most important and most widely studied technology in the field of Computational Advertising.Therefore,it is a very important and meaningful issue to use the large-scale historical advertising log to predict the advertising click-through rate effectively.Logistic regression has been widely used in advertising click-through rate prediction.However,the traditional logical regression has low accuracy and poor timeliness and so on in dealing with large-scale complex data sets.This paper first analyzes the advertising click log and prepares the data set according to the characteristics of the data set.On this basis,extracting six different types of features according to the advertising attributes,user attributes,site attributes.Afterwards,this paper uses the Logistic Regression model,the Naive Bayesian model,the Support Vector Regression model to predict the advertising click rate,and makes a comparative study of these three models.Then exploring the improvement of the model in the influence of the word vector feature and the high influence feature.The experimental results show that the word vector feature and the high influence feature make prediction results higher.Finally,due to the sparseness of the data,the large scale of the data and the slow training speed in the logistic regression model,The mixed regularization term is introduced in the model to prevent the training from fitting and the model is solved using the online optimization SGD and FTRL algorithm to improve the training efficiency.Then,the experimental results were analyzed by AUC and log-loss.The experimental results show that the improved logic regression model has higher accuracy and better timeliness.