KS-GNN: Keywords Search over Incomplete Graphs via Graphs Neural Network University of New South Wales University of New South Wales NSW, Australia

For PCA-based methods, the dimensionality reduction is performed via singular value decomposition (SVD) of the input one-hot encoding matrix X. As mentioned above, we utilize grid search for tuning the hyper-parameters. In particular, for the learning-based methods, including GraphSAGE and KS-GNN, the learning rates are selected from {0.1, 0.01, 0.001, 0.0001}. GraphSAGE, SAT, Conv-PCA, KS-PCA, KS-GNN), we swept the number of hidden layers in the set {1, 2, 3, 4, 5}. As for the margin hyper-parameter m in Eq.(6), we search it from {0, 0.1, 0.5, 1, 2.5, 5, 10}.