Structure preserved ordinal unsupervised domain adaptation

Unsupervised domain adaptation (UDA) aims to transfer the knowledge from labeled source domain to unlabeled target domain. The main challenge of UDA stems from the domain shift between the source and target domains. Currently, in the discrete classification problems, most existing UDA methods usually adopt the distribution alignment strategy while enforcing unstable instances to pass through the low-density areas. However, the scenario of ordinal regression (OR) is rarely researched in UDA, and the traditional UDA methods cannot preferably handle OR since they do not preserve the order relationships in data labels, like in human age estimation. To address this issue, we proposed a structure-oriented adaptation strategy, namely, structure preserved ordinal unsupervised domain adaptation (SPODA). More specifically, on one hand, the global structure information was modeled and embedded into an auto-encoder framework via a low-rank transferred structure matrix. On the other hand, the local structure information was preserved through a weighted pair-wise strategy in the latent space. Guided by both the local and global structure information, a well-performance latent space was generated, whose geometric structure was adopted to further obtain a more discriminant ordinal regressor. To further enhance its generalization, a counterpart of SPODA with deep architecture was developed. Finally, extensive experiments indicated that in addressing the OR problem, SPODA was more effective and advanced than existing related domain adaptation methods.