Domain Adaptation Algorithm based on Manifold Regularization

Abstract

Domain adaptation for classification is often encountered in recent years. A popular approach consists in transforming the source and target data to an identical linear space. Then the Maximum Mean Discrepancy (MMD) is used to evaluate the dissimilarity of distributions. However, the MMD only makes the source and target domain distribution consistent according to the global probability distribution, and cannot effectively protect the local geometric structure of the data. To make better use of the structure of local geometry, this paper proposes a method called domain adaptation based on manifold regularization (DAMR). First, this algorithm embeds the input data into a reproducing kernel Hilbert space (RKHS). Second, subspace-based dimensionality reduction is conducted on the RKHS. Third, a manifold regularization term is added to the learning method. Furthermore, the classification experiments demonstrate that DAMR is an accurate and effective method.