Drift-Aware Ensemble Regression

Abstract. Regression models are often required for controlling production processes by predicting parameter values. However, the implicit assumption of standard regression techniques that the data set used for parameter estimation comes from a stationary joint distribution may not hold in this context because manufacturing processes are subject to physical changes like wear and aging, denoted as process drift. This can cause the estimated model to deviate significantly from the current state of the modeled system. In this paper, we discuss the problem of estimating regression models from drifting processes and we present ensemble regression, an approach that maintains a set of regression models—estimated from different ranges of the data set—according to their predictive performance. We extensively evaluate our approach on synthetic and real-world data. Key words: Ensemble Method, Regression, Process Drift, Stochastic Process

• Type: Conference
• Year: 2009
• Where: MLDM
• Authors: Frank Rosenthal, Peter Benjamin Volk, Martin Hahmann, Dirk Habich, Wolfgang Lehner
• Keywords: Machine Learning, MLDM 2009, Process Drift, Regression Models, Standard Regression Techniques