Two-Step Projection of Sparse Discrimination Between Classes for Unsupervised Domain Adaptation

In the past few years, researchers have developed domain adaptive (DA) techniques which aim to address the domain shift between training and testing sets. However, most existing unsupervised domain adaptation techniques only use a projection matrix to train the classifier, which do not impose any restrictions on the same class samples. In this paper, we introduce a novel approach to unsupervised domain adaptation, called two-step projection of sparse discrimination between classes for unsupervised domain adaptation (TSPSDC). Unlike existing methods that use a single matrix for classifier learning, TSPSDC leverages two-step projection learning and integrates inter-class sparsity constraints to extract domain-invariant features from the class level. Specifically, 1) aiming to mitigate any adverse consequences arising from domain shift, distribution alignments are implemented to decrease the distribution disparity between the source and target domains in the shared subspace. 2) Our approach involves integrating two types of regularization: manifold regularization and inter-class sparsity regularization. The feature representation exhibits the same row sparsity structure within each class. Simultaneously, the distance between similar classes is minimized as a result. The resulting feature representation is thus advantageous for achieving highly discriminative representations. Extensive experimental results on three different sets of data demonstrate that the proposed approach outperforms many contemporary techniques for domain adaptation.