Bootstrapping and Bagging Data with Multi-Way Dependencies

• Activity Number: 328
• Type: Topic Contributed
• Date/Time: Tuesday, July 31, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Dean Eckles*+ and Art B Owen
• Companies: Facebook and Stanford University
• Keywords: online bootstrap ; online bagging ; unbalanced random effects ; advertising ; ensemble methods ; transposable data

Abstract

Computerized data gathering frequently produces data sets with two-way, three-way, or even higher order data tables. We present a bootstrap for such tensor valued data. Our version uses independent weights instead of multinomial ones. It is mildly conservative. Poisson weights are close to the original bootstrap, but binary weights have computational and statistical advantages. Under certain conditions a single bootstrap replicate suffices to give a variance estimate. We apply our method to (1) experiments identifying peer effects in social advertising and (2) comparisons of comment lengths by British and American Facebook users via Web and mobile interfaces. This method can also be used in bootstrap aggregation (bagging) for training with dependent data.