Source-free unsupervised domain adaptation (SFDA), which needs only pre-trained source model and unlabeled target data ( \({D}_{t}\) ), has attracted lots of interests recently.Query Compared with unsupervised domain adaptation, SFDA is more practical since it can be applied to scenarios that lack of source data ( \({D}_{s}\) ) due to data privacy. Although remarkable progresses have been made in the field of SFDA, it still remains challenge since most of the existing methods do not fully exploit the internal structure information of the \({D}_{t}\) . To bridge the gap, we propose prototype generative model leveraging inherent data structure mined by graph neural networks and forge a source-free domain adaptation method accordingly. Our method consists of two stages: (1) data mining and prototype generation and (2) domain adaptation. In stage one, we firstly use graph convolutional network and graph autoencoder to model the data relations of the \({D}_{t}\) , extract features and label the target features in a self-supervised way; then, we generate source-like avatar conditioned on features of the \({D}_{t}\) , labels and noise. In the second stage, a category-level alignment between the source prototypes and the \({D}_{t}\) is conducted based on contrastive learning strategy. Extensive experiments on Office-31, Office-Home and VisDA-2017 datasets prove the effectiveness and superiority of our method. Average accuracy rates of 90.7%, 76.1% and 87.4% are achieved, thereby leading to superior performance compared with the state-of-the-art methods.
