Modified Covariate-shift Multi-source Ensemble Method in Transferability Metric

Transfer learning usually focuses on dealing with small training set in target domain by sharing knowledge generated from source ones, in which one main challenge is divergence metric of distributed samples between training and test data. In order to deal with negative problem caused by improper auxiliary sample selections in source domains, this paper presents a modified covariate-shift multi-source ensemble method with transferability criterion. Firstly, transferability metric of auxiliary samples is defined by joint density estimation in accordance with co-variant transfer principles from source to target, so that the coherency of data distributions is verified. After that, whether transfer learning occurs or not should be determined after evaluating transferability metric in different sources to boost accuracy. Finally, experiments on Caltech256 using GIST demonstrate effectiveness and efficiency in the proposed approach and discussions of performance under diverse selections from auxiliary samples and source domains are presented as well. Experimental results show that the proposed method can sufficiently hold back negative for better learnability in transfer style.