Anti-Symmetric DGN: a stable architecture for Deep Graph Networks

Authors: Alessio Gravina, Davide Bacciu, Claudio Gallicchio

ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically validate the proposed approach on several graph benchmarks, showing that A-DGN leads to improved performance and enables to learn effectively even when dozens of layers are used. We conduct extensive experiments to demonstrate the benefits of our method.
Researcher Affiliation Academia Alessio Gravina University of Pisa alessio.gravina@phd.unipi.it Davide Bacciu University of Pisa davide.bacciu@unipi.it Claudio Gallicchio University of Pisa claudio.gallicchio@unipi.it
Pseudocode No The paper provides mathematical equations and descriptions of the model, but it does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes We release openly the code implementing our methodology and reproducing our empirical analysis at https://github.com/gravins/Anti-Symmetric DGN.
Open Datasets Yes For the graph property prediction task, we considered three datasets extracted from the work of Corso et al. (2020). In the graph benchmark setting we consider five well-known graph datasets for node classification, i.e., Pub Med (Namata et al., 2012); coauthor graphs CS and Physics; and the Amazon co-purchasing graphs Computer and Photo from Shchur et al. (2018).
Dataset Splits Yes As in the original work, we used 5120 graphs as training set, 640 as validation set, and 1280 as test set. Similarly to Shchur et al. (2018), we use 20 labeled nodes per class as the training set, 30 nodes per class as the validation set, and the rest as the test set.
Hardware Specification Yes We carried the experiments on a Dell server with 4 Nvidia GPUs A100.
Software Dependencies No The paper mentions software components like 'Adam optimizer' and 'Multi-Layer Perceptrons' but does not provide specific version numbers for any programming languages or libraries used in the implementation.
Experiment Setup Yes We performed hyper-parameter tuning via grid search, optimizing the Mean Square Error (MSE). We trained the models using Adam optimizer for a maximum of 1500 epochs and early stopping with patience of 100 epochs on the validation error. We report in Appendix F the grid of hyper-parameters exploited for this experiment.