Reliable Graph Neural Networks via Robust Aggregation

Authors: Simon Geisler, Daniel Zügner, Stephan Günnemann

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, our method improves the robustness of its base architecture w.r.t. structural perturbations by up to 550% (relative), and outperforms previous state-of-the-art defenses.
Researcher Affiliation Academia Simon Geisler Daniel Zügner Stephan Günnemann Department of Informatics Technical University of Munich {geisler, zuegnerd, guennemann}@in.tum.de
Pseudocode No The paper describes algorithms and methods verbally and through equations but does not include structured pseudocode or algorithm blocks.
Open Source Code Yes The source code is available at https://www.daml.in.tum.de/reliable_gnn_via_robust_aggregation.
Open Datasets Yes We evaluate these models on Cora ML [47], Citeseer [41], and Pub Med [47] for semisupervised node classification.
Dataset Splits Yes For each approach and dataset, we rerun the experiment with three different seeds, use each 20 labels per class for training and validation, and report the one-sigma error of the mean.
Hardware Specification Yes We used one 2.20 GHz core and one Ge Force GTX 1080 Ti (11 Gb).
Software Dependencies No The paper mentions general tools or frameworks (e.g., GNNs, Deep Robust's implementation) but does not provide specific version numbers for software libraries or dependencies.
Experiment Setup Yes We set the number of hidden units for all architectures to 64, the learning rate to 0.01, weight decay to 5e 4, and train for 3000 epochs with a patience of 300. For the architectures incorporating our Soft Medoid, we perform a grid search over different temperatures T... In the experiments on Cora ML and Citeseer we use = 0.15 as well as k = 64. We use = 0.15 as well as k = 32 in the Pub Med experiments.