Pruning Ensemble Models for Classification Problems on Large Data Sets

• Activity Number: 236
• Type: Contributed
• Date/Time: Monday, August 1, 2011 : 2:00 PM to 3:50 PM
• Sponsor: Section on Statistical Computing
• Author(s): Damir Spisic*+ and Fan Li and Jing Xu
• Companies: IBM Research and IBM Research China and IBM Research China
• Keywords: classification ; ensemble models ; ensemble pruning ; large data modeling

Abstract

Classification methods that generate a single model are generally not suitable for analysis of very large datasets, streaming data or updated data. We consider an ensemble modeling approach where component models are generated on disjoint consecutive data blocks. This approach is efficient and effectively addresses all three stated challenges. Several learning methods are used for creating component models and a few different combination methods are considered for computing ensemble predictions. Ensembles are pruned in order to keep the number of component models limited and therefore efficient for model scoring. We study and compare several pruning methods in this context using validation and test samples extracted from the overall dataset. Our analysis is based on the accuracy measure. We assume stationary and randomly ordered datasets and show that this approach is not only efficient and scalable, but also produces similar or superior prediction accuracy when compared to the single model generated by the same learning method.