Improving Cost Estimation in Internet Advertising Using Machine Learning: Preliminary Results

Abstract

In the internet advertising industry, web and mobile applications that display ads need to choose high-paying ads to increase their revenue. Ad mediators create various decision mechanisms to select ads that will generate higher revenues in order to increase the revenue of advertising applications. One type of these decision mechanisms is to select and deliver the ad with the highest eCPM (Effective Cost Per Mille) value from ads that can be placed in an ad slot. The eCPM value varies depending on different external factors for different applications. It is not possible for domain experts to make successful predictions by analyzing different sets of external factors for many applications and to keep these predictions constantly updated. Therefore, eCPM values were automatically predicted separately for each application on different ad slots and different countries using time series analysis and machine learning algorithms. SARIMA, MLP, CNN and LSTM algorithms are used to make predictions. The LSTM algorithm has generally yielded better results in eCPM estimation. As a result of the trials conducted with a limited number of users of the two applications on production environment, an increase in daily income per user was observed.