Unsupervised Domain Adaptation based on the Geography Structural Information

Abstract

Traditional domain adaptation aligns the marginal and conditional distribution of the source and target domain. While the representation after projection is not discriminative enough for the final classification. In this paper, we propose a new method, the conditional distribution of both domains is aligned, the class information is employed to further enhance the discriminability of the samples. After aligning the class information, samples can be classified easier. Meanwhile, the geographic structure information in the samples is well preserved to train the classifier of the model. Several experiments are done to prove the effectiveness of the model and demonstrated good performance on three frequently used data sets: Office-Home, Amazon-Review, PIE.