Building powerful and equivariant graph neural networks with structural message-passing

Authors: Clément Vignac, Andreas Loukas, Pascal Frossard

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimentally, our model can predict various graph topological properties on synthetic data more accurately than previous methods and achieves state-of-the-art results on molecular graph regression on the ZINC dataset.
Researcher Affiliation Academia Clément Vignac, Andreas Loukas, and Pascal Frossard EPFL Lausanne, Switzerland {clement.vignac,andreas.loukas,pascal.frossard}@epfl.ch
Pseudocode No The paper describes its methods using mathematical equations and textual descriptions but does not include any clearly labeled pseudocode or algorithm blocks.
Open Source Code Yes Our implentation with Pytorch Geometric [51] is available at github.com/cvignac/SMP.
Open Datasets Yes The ZINC database is a large scale dataset containing molecules with up to 37 atoms. The task is to predict the constrained solubility of each molecule, which can be seen as a graph regression problem. We follow the setting of [55]: we train SMP on the same subset of 10,000 molecules with a parameter budget of around 100k, reduce the learning rate when validation loss stagnates, and stop training when it reaches a predefined value.
Dataset Splits Yes Models are retrained for each cycle length and graph size on 10k samples with balanced classes, and evaluated on 10, 000 samples as well. (Section 5.1) The training set is composed of 5120 graphs with up to 24 nodes, while graphs in the test set have up to 19 nodes. (Section 5.2) For ZINC: reduce the learning rate when validation loss stagnates (Section 5.3)
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU/CPU models, memory) used for running its experiments.
Software Dependencies No The paper mentions using 'Pytorch Geometric' but does not provide specific version numbers for this or any other software dependencies.
Experiment Setup Yes Models are retrained for each cycle length and graph size on 10k samples with balanced classes... (Section 5.1). For ZINC: We use an expressive parametrization of SMP, with 12 layers and 2-layer MLPs both in the message and the update functions. In order to reduce the number of parameters, we share the same feature extractor after each layer (cf Fig. 2). We follow the setting of [55]: we train SMP on the same subset of 10,000 molecules with a parameter budget of around 100k, reduce the learning rate when validation loss stagnates, and stop training when it reaches a predefined value. (Section 5.3)