An Empirical Evaluation of Ensemble Decision Trees to Improve Personalization on Advertisement

Ensemble decision tree algorithms are well known for good prediction accuracy in most cases, but not much research has been done on applying ensemble methods to improve personalization in the field of behavioral targeting in online advertisements. In behavioral targeting, the best ad is matched to the user based on his/her past activities and demographics. At present, most models used in the behavioral targeting are some form of linear models. Our goal in this paper is to analyze and understand the effect of ensemble techniques on large scale advertising data. Few of the main challenges of this kind of large scale data are sparse features and high dimensionality that make it hard for one single model to work the best. The form of ensemble method explored in this paper is the random forest based regression algorithm that combines the power of multiple decision trees to produce a more robust model which has a reduced variance as well bias. Also, in the field of online advertising it is imperative to learn in an online fashion (while the advertising campaign is being run) as the customers want to get the most off their money at the earliest and the lifetime of such advertisements is short. So, some form of exploration vs. exploitation technique is also required to be used in the system. Our contributions in this paper are three fold. First, we develop a new technique to determine optimal parameters of the random forest algorithms. Second, we do a comparative analysis of random forest vs. logistic regression. Third, we combine ensemble decision tree algorithms with bandit algorithms to produce around 17% CTR improvement over random.