Research On High Level Feature Representation And Predicting Methods In Online Advertising

With the rapid development of advertising industry, the revenue of the ads hasincreasingly become the main source of income for many companies. In order to betteroptimize advertising effect to obtain greater benefits and make up for some shortcomingsof traditional advertising, the online advertising rises rapidly, using computationalmethods to improve the efficiency of advertising. Currently, the sponsored searchadvertising and real-time bidding advertising are two important forms of onlineadvertising. Since the cutting-edge research and development in computationaladvertising have connected more and more with the Big Data, the research communityhas draw a lot of attention on some related issues, among them, predicting the click-through rate (CTR) is one of the hottest topics. Therefore how to use sophisticatedadvertising logs to estimate the CTR accurately has also been an extremely important job.In this paper, our target is predicting the CTR of sponsored search advertising andreal-time bidding advertising more accurately, we gain high level features based on thebasic features through deep network framework we proposed, and then combine the otherbasic features and the high level features for a better predicting. Specifically, the studymainly includes the following aspects.Firstly, we analyze the complex fields of the advertising logs and do related workfor dataset pre-processing. Then introduce the popular evaluation metris of click-throughrate prediction. On this basis, utilize topic model, similarity algorithms and statisticalinformation to extract some basic features including categorical feature, historical CTRinformation, similarity feature, interesting and tag information and hash value feature,and analyze the effectiveness of them.Secondly, in this paper, we use na ve bayes model and support vector regressionmodel to predict click-through rate. Based on the characteristics of each model, we selectcategorical features into na ve bayes model to estimate CTR, and select historicalinformation and behavioral features into support vector regression model to predict. Thenaccording to the predicting result, get more expressive features for furture research. Inaddition, considering the limtations and performance in different feature sets of the singlemodel, we propose an ensenmble method to predict CTR.Finally, due largely to the limited expression performance of the shallow basicfeatures, they can hardly mine their relationship and latent factors. So after a detailedanalysis of the raw basic features, we propose a deep network to help predict CTR. Forthe features selected in previous chapters, we put them into different structures of deepnetwork according to the attribute of themselves to learn high level features, after gettingthem, we concatenate the high level feature and the other basic feature, and then use logistic regression model and support vector regression model to predict the click-throughrate. We conduct experiments on sponsored search advertising dataset and real-timebidding advertising dataset. The experimental results demonstrate the effectiveness of thehigh level features and our predicting method can perform well on predicting CTR.