In this paper, we tackle a new problem of \textit{multi-source unsupervised
domain adaptation (MSUDA) for graphs}, where models trained on annotated source
domains need to be transferred to the unsupervised target graph for node
classification. Due to the discrepancy in distribution ac