Bridged Refinement for Transfer Learning

Abstract

There is usually an assumption in traditional machine learning that the training and test data are governed by the same distribution. This assumption might be violated when the training and test data come from different time periods or domains. In such situations, traditional machine learning methods not aware of the shift of distribution may fail. This paper proposes a novel algorithm, namely bridged refinement, to take the shift into consideration. The algorithm corrects the labels predicted by a shift-unaware classifier towards a target distribution and takes the mixture distribution of the training and test data as a bridge to better transfer from the training data to the test data. In the experiments, our algorithm successfully refines the classification labels predicted by three state-of-the-art algorithms: the Support Vector Machine, the naïve Bayes classifier and the Transductive Support Vector Machine on eleven data sets. The relative reduction of error rates is about 50% in average.