Discriminative transfer feature learning for unsupervised domain adaptation

Feature extraction-based methods have been widely applied in unsupervised domain adaptation. Recent advances focus on extracting transferable features that are not only domain-variant but also category-discriminative. However, the ordinal locality of data has been scarcely explored in the construction of discriminative terms. In this paper, we propose a novel unsupervised domain adaptation method called discriminative transfer feature learning (DTFL) for pattern classification. Specifically, DTFL first constructs a class-aware ordinal locality-preserving term (COLP). COLP helps preserve the inherent neighborhood relationships of each training sample and inherit the geometric structure information of training samples, thereby enhancing the class-wise discriminability of the learned transferable features. Then, label consistency within domains is used to refine the target pseudo-label. Moreover, the discriminative transfer feature learning and refined target labels can mutually benefit from each other iteratively. Experiments conducted on three cross-domain visual datasets and two traditional intrusion detection datasets demonstrate that the proposed method exceeds existing methods in classification performance.