Knowledge Management and Acquisition for Intelligent Systems

Abstract

Data shifting in machine learning problems violates the common assumption that the training and testing samples should be drawn from the same distribution. Most of the algorithms which provide the solution for data shifting problems first try to evaluate the distributions and then reweight samples based on their distributions. Due to the difficulty of evaluating a precise distribution, conventional methods cannot achieve good classification performance. In this paper, we introduce two types of data-shift problems and propose a model-based co-clustering transfer learning based solution which consistently deals with both scenarios of data shift. Experimental results demonstrate that our proposed method achieves better generalization and running efficiency compared to tra-ditional methods under data or covariate shift setting.