Comparison of R and Vowpal Wabbit for Click Prediction in Display Advertising

• Activity Number: 41
• Type: Contributed
• Date/Time: Sunday, August 9, 2015 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Jaimyoung Kwon* and Bin Ren and Rajasekhar Cherukuri and Marius Holtan
• Companies: AOL Advertising and AOL Platforms and AOL Platforms and AOL Platforms
• Keywords: click prediction ; display advertising ; lasso ; R ; Vowpal Wabbit ; Hadoop

Abstract

Efficient operation of online display advertising campaigns relies on accurate prediction of the likelihood that a web user clicks on a particular banner on a particular site. The problem, which is one of canonical examples of the application of machine learning and statistics to the "big data", is essentially a classification problem with many predictor variables, where petabytes of input data are processed in a Hadoop pipeline. Commonly applied methods include lasso, random forest, and gradient boosting, implemented on such computing platforms as R, Vowpal Wabbit, and native implementation. Among these, Vowpal Wabbit, an out-of-core learning system, has become a popular choice in many large scale machine learning problems. In this paper, we compare the performance of a few different algorithms and platform combinations. In particular, we compare the performance of algorithms implemented in R and Vowpal Wabbit in various sample size and feature number combinations, and their interactions with sampling, feature selection, and feature transformation.