Application Research Of Prediction Algorithms In Online Advertising

Online advertising is one of the Internet’s most profitable businesses. As a kind of online advertising, display advertising has been widely studied by researchers and engineers especially for Guaranteed Display Advertising (GD). In guaranteed display advertising, advertisers and advertising platforms will reach an agreement to ensure that advertising platforms will display the advertisement a specified number of times. According to the contract, advertisers can purchase a certain number of user visits from the Internet publisher several months in advance, and the publisher need to guarantee these visits. The publisher will get paid if they meet the advertiser’s requirements, otherwise they will incur penalties if the guarantees are not met. Page visits forecasting for a website becomes a vital part of display advertising, and the accuracy of forecasting will influence the revenue of advertising platform directly. To achieve the goal, we need to collect history data of page visits, and analyze it by using some time series forecasting methods. Advertising platform will take the risk of compensating for breaking a contract if we do an excessive prediction. On the other hand, it will lead to reduction in income if we take a conservative estimate of page visits. Therefore, how to analyze the history page visits data in a better way and improve the accuracy of prediction has become an important problem that needs to be solved.In this dissertation we try to solve the problem of page visits forecasting for a website by treating the history page visits data as a time series data, and then exploit time series forecasting models to analyze the history data. The main works of the dissertation are listed here. We study the performance of ARIMA model and Holt-Winters seasonal model on page visits forecasting respectively. Since Holt-Winters seasonal model has two kinds of seasonality:multiplicative and additive, we can divide Holt-Winters method into multiplicative Holt-Winters seasonal model and additive Holt-Winters seasonal model. This dissertation analyzes the performances of ARIMA model, multiplicative Holt-Winters method and additive Holt-Winters method in a real world dataset by comparing with each other. Experimental results show that the performances of ARIMA method and multiplicative Holt-Winters method outperform additive Holt-Winters method in most cases. The experimental results also indicate that ARIMA method and multiplicative Holt-Winters method are suitable for different scenarios. ARIMA method is applicable to time series data which doesn’t have a significant cyclical, while multiplicative Holt-Winters method achieves the best performance for periodic time series data. In order to improve the accuracy of page visits forecasting and filter outliers, we apply traditional filtering methods, including frequency filtering method, Mean filtering method and Gaussian filtering method, to preprocess the time series data. The results indicate that the performances of traditional filtering algorithms are not good. This dissertation has proposed a traffic splitting method which is based on sliding windows to solve the problem, and compared it with traditional filtering methods. Experimental results show that the proposed filtering algorithm has a greater improvement on the accuracy of page visits forecasting.