Hierarchical semisupervised transfer AdaBoost

We propose a hierarchical and semisupervised transfer AdaBoost (HissTrAdaBoost) algorithm to address over-fitting and generalization problem in TrAdaBoost, which is one of the state-of-the-art instance based transfer learning algorithm. Specifically, the samples in the source domain which have larger difference from the target domain are removed, and then the unlabeled instances in the target domain are hierarchically imported to the classifiers. In this way, the generalization error is reduced by extra constraints provided by the semi-supervised classifiers of the unlabeled data. Experimental results conducted on the public data sets confirm the effectiveness of the proposed method, for the classification accuracy has been improved by 1% to 3%.