Open Set Domain Adaptation with Zero-shot Learning on Graph

Open set domain adaptation focuses on transferring the information from a richly labeled domain called \emph{source domain} to a scarcely labeled domain called \emph{target domain} while classifying the unseen target samples as one \emph{unknown} class in an unsupervised way. Compared with the close set domain adaptation, where the source domain and the target domain share the same class space, the classification of the unknown class makes it easy to adapt to the realistic environment. Particularly, after the recognition of the unknown samples, the robot can either ask for manually labeling or further develop the classification ability of the unknown classes based on pre-stored knowledge. Inspired by this idea, in this paper we propose a model for open set domain adaptation with zero-shot learning on the unknown classes. We utilize adversarial learning to align the two domains while rejecting the unknown classes. Then the knowledge graph is introduced to generate the classifiers for the unknown classes with the employment of the graph convolution network (GCN). Thus the classification ability of the source domain is transferred to the target domain and the model can distinguish the unknown classes with prior knowledge. We evaluate our model on digits datasets and the result shows superior performance.