Discriminative Noise Robust Sparse Orthogonal Label Regression-Based Domain Adaptation

Domain adaptation ( DA ) aims to enable a learning model trained from a source domain to generalize well on a target domain, despite the mismatch of data distributions between the two domains. State-of-the-art DA methods have so far focused on the search of a latent shared feature space where source and target domain data can be aligned either statistically and/or geometrically. In this paper, we propose a novel unsupervised DA method, namely D iscriminative Noise Robust Sparse O rthogonal L abe l Regression-based D omain A daptation ( DOLL-DA ). The proposed DOLL-DA derives from a novel integrated model which searches a shared feature subspace where data labels are orthogonally regressed using a label embedding trick, and source and target domain data are discriminatively aligned statistically through optimization of some repulse force terms. Furthermore, in minimizing a novel Noise Robust Sparse Orthogonal Label Regression (NRS_OLR) term, the proposed model explicitly accounts for data outliers to avoid negative transfer and introduces the property of sparsity when regressing data labels. We carry out comprehensive experiments in comparison with 35 state of the art DA methods using 8 standard DA benchmarks and 49 cross-domain image classification tasks. The proposed DA method demonstrates its effectiveness and consistently outperforms the state-of-the-art DA methods with a margin which reaches 17 points on the CMU PIE dataset. To gain insight into the proposed DOLL-DA , we also derive three additional DA methods based on three partial models from the full model, namely OLR , CDDA+ , and JOLR-DA , highlighting the added value of (1) discriminative statistical data alignment; (2) Noise Robust Sparse Orthogonal Label Regression; and (3) their joint optimization through the full DA model. In addition, we also perform time complexity and an in-depth empiric analysis of the proposed DA method in terms of its sensitivity w.r.t. hyper-parameters, convergence speed, impact of the base classifier and random label initialization as well as performance stability w.r.t. target domain data being used in training.