Visual Domains Adaptation via Feature and Model Matching

In most machine learning algorithms, the distribution of training and test sets (source and target domains, respectively) are assumed the same. However, this condition is violated in many real world problems and the performance of model degrades as well. The aim of domain adaptation solution is to build an adaptive model on source data to have acceptable performance on target domain. In this paper, we propose an unsupervised two-phases approach which benefits from representation and model adaptation methods. In the first phase, source and target data are projected onto a common subspace on which the marginal and conditional distribution difference is minimized. Moreover, domain invariant clustering is exploited to discriminate between various classes of source data. In the second phase, an adaptation classifier is presented to minimize prediction error rate and maximize manifold adaptability across source and target domains. The proposed approach is evaluated on four visual benchmark datasets according to 36 designed experiments. The obtained results highlight the considerable performance of the proposed approach against other state-of-the-art machine learning and transfer learning methods.