Open Set Domain Adaptation: Theoretical Bound and Algorithm

The aim of unsupervised domain adaptation is to leverage the knowledge in a labeled (source) improve model’s learning performance with an unlabeled (target) domain—the basic strategy being mitigate effects discrepancies between two distributions. Most existing algorithms can only handle closed set (UCSDA), i.e., where source and target domains are assumed share same label set. In this article, we more challenging but realistic setting: open (UOSDA), has unknown classes that not found domain. This first study provide bound for adaptation, which do by theoretically investigating risk classifier on classes. proposed special term, namely, difference, reflects Furthermore, present novel guided algorithm called distribution alignment difference (DAOD), based regularizing difference bound. experiments several benchmark data sets show superior UOSDA method compared state-of-the-art methods literature.