Robust prediction for quality of industrial processes

Abstract

This paper proposes a new robust predictive approach for quality of industrial processes. It draws inspiration from robust AdaBoost for classification and expands to regression tasks. Existing classical AdaBoost for regression (AdaBoost.R2) constructs a strong learner in a stepwise fashion by re-weighting those instances according to their regression results at each iteration. In order to reduce its sensitivity to outliers, the proposed approach shows how the weight can be modified by a mixture of exponential updates with additional uniform weight for predictive problems. Experimental results using actual data from an ore-dressing production processes show its more robustness than existing methods even if a certain amount of data is infected.