Effects of drift and noise on the optimal sliding window size for data stream regression models. Issue 10 (19th May 2017)

• Record Type: Journal Article
• Authors: Tschumitschew, Katharina; Klawonn, Frank
• Is Part Of: Communications in statistics. Volume 46:Issue 10(2017)
• Page Start: 5109
• Page End: 5132
• Publication Date: 2017-05-19
• Subjects: Adaptive regression -- concept drift -- data stream analysis -- online learning.; 68T05 -- 68Q32 -- 91E40; Mathematical statistics -- Periodicals; Mathematics; Statistics; 519.2
• DOI: 10.1080/03610926.2015.1096388
• Languages: English
• ISSNs: 0361-0926

Abstract

ABSTRACT: The analysis of non stationary data streams requires a continuous adaption of the model to the relevant most recent data. This requires that changes in the data stream must be distinguished from noise. Many approaches are based on heuristic adaptation schemes. We analyze simple regression models to understand the joint effects of noise and concept drift and derive the optimal sliding window size for the regression models. Our theoretical analysis and simulations show that a near optimal window size can be crucial. Our models can be used as benchmarks for other models to see how they cope with noise and drift.