Research And Design Of Display Advertising System Based On Online Learning

With the rapid development of the Internet,the status of online advertising,which is one of the most important means of revenue in Internet industry,is becoming increasingly important.CTR prediction of online advertising is an important research topic in the Internet.Its results are closely related to user attributes,advertisement attributes and contextual information.The result of CTR prediction is very important to improve the income of Internet companies.This thesis focuses on display advertising,introduces and analyzes the structure of the online advertising system,and expounds the important position of the CTR prediction in the advertising system.This thesis focuses on the following three problems in display advertising system;The first problem is about constructing feature engineering platform.We use big data technology to process data and extract useful features in both offline and real-time manners,and load the results into distributed database.Secondly,the logistic regression(LR),which is the most common model for CTR prediction,is essentially a generalized linear model.This thesis considers user features and advertising features,thus constructing the relationship between advertising features and user features.In this way we can improve the accuracy of CTR prediction combined with the optimization of the traditional training model.Thirdly,it takes too long for traditional model of offline training to update the model itself.Based on this,this thesis presents the online learning training methods.We make multiple experimental comparisons of FTRL online learning algorithm and the traditional LR algorithm.And it proves that online learning can update the model and response to the changes online very quickly.And the online learning algorithm can improve the click-through rate online,thus increase the income for companies and enhance the interests of all parties in online advertising.