This is the appendix of our work: 'GNNEvaluator: Evaluating GNN Performance On Unseen We provide the details of dataset statistics used in our experiments in Table. For all GNN and MLP models, the default settings are: (a) the number of layers is 2; (b) the hidden feature dimension is 128; (c) the output feature dimension before the softmax operation is 16. The hyperparameters of training these GNNs and MLP are shown in Table A2. As a vital component of our proposed two-stage GNN model evaluation framework, DiscGraph set captures wide-range and diverse graph data distribution discrepancies. In Fig. A1, we present more visualization results on discrepancy node attributes in the proposed DiscGraph set for different GNN models, i.e., (a) GA T, (b) GraphSAGE, and (c) GIN, under

Theoretically, we show that it is possible for our approach to simultaneously improve both accuracy and fairness--in contrast to standard fairness approaches that suffer from tradeoffs. Empirically, we show that our technique improves accuracy on weak supervision baselines by as much as 32% while reducing demographic parity gap by 82.5%.

The sample explanatory features were fed into a multi-layer perceptron, then the learned latent features and sample spatial locations were fed into a Gaussian process model. GP variance is used as the uncertainty measure. We first constructed a spatial graph based on each sample's k-nearest-neighbor by spatial distance. The model contains two GCN layers. It contains a multi-level graph neural network to capture the long-range interactions among particles with linear complexity.