There has been a recent surge of interest in designing Graph Neural Networks\n(GNNs) for semi-supervised learning tasks. Unfortunately this work has assumed\nthat the nodes labeled for use in training were selected uniformly at random\n(i.e. are an IID sample). However in many real world scenarios gathering labels\nfor graph nodes is both expensive and inherently biased -- so this assumption\ncan not be met. GNNs can suffer poor generalization when this occurs, by\noverfitting to superfluous regularities present in the training data. In this\nwork we present a method, Shift-Robust GNN (SR-GNN), designed to account for\ndistributional differences between biased training data and the graph's true\ninference distribution. SR-GNN adapts GNN models for the presence of\ndistributional shifts between the nodes which have had labels provided for\ntraining and the rest of the dataset. We illustrate the effectiveness of SR-GNN\nin a variety of experiments with biased training datasets on common GNN\nbenchmark datasets for semi-supervised learning, where we see that SR-GNN\noutperforms other GNN baselines by accuracy, eliminating at least (~40%) of the\nnegative effects introduced by biased training data. On the largest dataset we\nconsider, ogb-arxiv, we observe an 2% absolute improvement over the baseline\nand reduce 30% of the negative effects.\n