Graph Adversarial Self-Supervised Learning

Authors: Longqi Yang, Liangliang Zhang, Wenjing Yang

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental on ten graph classification datasets show that the proposed approach is superior to state-of-the-art self-supervised learning baselines, which are competitive with supervised models.
Researcher Affiliation Academia Longqi Yang Institute for Quantum Information & State Key Laboratory of High Performance Computing, College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, China Defense Innovation Institute, Beijing 100071, China ... Liangliang Zhang Institute of Systems Engineering, AMS, Beijing, China ... Wenjing Yang Institute for Quantum Information & State Key Laboratory of High Performance Computing, College of Computer Science and Technology, National University of Defense Technology, Changsha 410073, China
Pseudocode Yes Algorithm 1 Graph Adversarial Self-Supervised Learning (GASSL)
Open Source Code No The paper does not provide any explicit statements or links to open-source code for the described methodology.
Open Datasets Yes We selected 10 widely used graph classification datasets from TU datasets [26] and Open Graph Benchmark (OGB) [27]. ... For OGB datasets, we evaluate the performance with their original feature extraction and following the original training/validation/test dataset splits [27].
Dataset Splits Yes For all experiments on the TU dataset, we follow [17, 31] and report the mean 10-fold cross-validation accuracy with standard deviation after 5 runs followed by a linear SVM. ... For OGB datasets, we evaluate the performance with their original feature extraction and following the original training/validation/test dataset splits [27].
Hardware Specification Yes We conduct all the experiments on an Nvidia TITAN Xp.
Software Dependencies No The paper mentions using Adam optimizer, GCN, and GIN, but does not provide specific version numbers for any software libraries or dependencies.
Experiment Setup Yes We train the model using Adam optimizer with an initial learning rate of 10 4, and we choose the number of GCN and GIN layers {2, 3, 4, 5}, number of epochs {20, 40, 100, 200}, batch size {32, 64, 128, 256, 512, 1024}, and the SVM parameter C {10 3, 10 2, . . . , 102, 103}. The step size α is set to 8 10 3, the perturbation bound ϵ is set to 8 10 3, the embedding dimension is set to 128 (expect HIV set to 512). We also use early stopping with the patience of 20, where we stop training if there is no further improvement on the validation loss during 20 epochs.