A Research On CTR Optimization Of Online Advertising Based On Logistic Regression

Click-through Rate(CTR) prediction is an important research topic for internet companies. The results are closely related to the context, user attributes and advertising attributes, and the effective prediction of CTR is essential to improve the company’s revenue.The most common model for CTR prediction is logistic regression (LR), but LR is essentially a generalized linear model. For the calculation of advertising, there are not only many features, but also the relationship among features. On the one hand, the actual business needs to quickly and effectively screen features and feature combinations, which in a large extent can only rely on artificial experience, time-consuming and not necessarily bring about the effect of the upgrade. On the other hand, the advertising data is produced in real time, and the processing of batch will face the problem of timeliness, online computing is particularly important.Based on introduction of principles and parameters optimization algorithm for the traditional LR model.we get the user features and advertising features,then XTUS(XTaV)T indicates the relationship between the user and the advertisement, by adding it to the sigmoid function and obtain a new LR model. Different from the previous methods, this paper uses online optimization algorithm FTRL to improve the efficiency of the parameters, using the mixed regularization to prevent the training over fitting. In the experimental part, we consider two types of indicators, AUC and log-loss, based on the previous model and algorithm, we analyze the accuracy, sensitivity and reliability of the parameters and draw the conclusion.The main results of this paper have three points. First,we consider the characteristics of users and advertising in this paper, and get the decomposition of the correlation matrix based on the sparsity and large scale of the correlation matrix,which makes it has a stronger non-linear fitting ability than the traditional LR model.Second, the model of this paper can automatically eliminate the useless features, making online prediction more rapidly, especially for large scale sparse data and features.Third, the use of FTRL online algorithm can process data in real-time,which make its efficiency increase and avoid the batch processing.