A Semisupervised Domain Adaptation Framework Using Dynamic Distribution Alignment and Manifold Regularization

Domain adaptation (DA) aims to build classifiers for a target domain (TD) by utilising knowledge from a source domain (SD). However, due to two main issues: lack of training data in TD and distribution divergence between the domains, DA can be a highly challenging problem in reality, especially as the existing methods often only prioritise one issue at a time. In this paper, we present a method called SDM that builds a classifier for TD having some unlabelled data by leveraging knowledge from SD having huge labelled data. We first build a classifier applying an existing classification algorithm, such as SVM, on the training data having only SD data. The classifier is then iteratively adjusted by updating the training data, adding TD data and deleting SD data in each iteration. We alleviate both issues by simultaneously minimising the structural risk and joint distributions across domains, and maximising the manifold consistency underneath marginal distributions. Furthermore, while existing methods rely on a user-defined threshold for adding and deleting data, we propose an automatic strategy to determine them. We evaluate the effectiveness of SDM by comparing its performances with fifteen state-of-the-art methods on seven publicly available datasets. We also assess the performance of SDM under certain challenging conditions, such as SD and TD datasets may have noisy and missing values. Our experimental results, including statistical significance tests, demonstrate a clear superiority of SDM over the state-of-the-art methods used in this study. SDM’s source code is publicly available on GitHub athttps://github.com/grahman20/SDM.