Weisfeiler and Lehman Go Cellular: CW Networks

Authors: Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yuguang Wang, Pietro Liò, Guido F. Montufar, Michael Bronstein

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this section we validate the theoretical and empirical properties of our proposed message passing scheme in controlled scenarios as well as in real-world graph classification problems, with a focus on large scale molecular benchmarks. Table 1: Classification accuracy on CSL. Table 2: TUDatasets. Table 3: ZINC (MAE), ZINC-FULL (MAE) and Mol-HIV (ROC-AUC).
Researcher Affiliation Collaboration Cristian Bodnar University of Cambridge cb2015@cam.ac.uk Fabrizio Frasca Imperial College London & Twitter ffrasca@twitter.com UCLA otter@math.ucla.edu Yu Guang Wang MPI-MIS, SJTU & UNSW yuguang.wang@unsw.edu.au Pietro Liò University of Cambridge pl219@cam.ac.uk Guido Montúfar MPI-MIS & UCLA montufar@math.ucla.edu Michael Bronstein Imperial College London & Twitter mbronstein@twitter.com
Pseudocode No The paper describes the steps of the CWL algorithm in a numbered list (e.g., '1. Given a regular cell complex X, all the cells σ are initialised with the same colour. 2. Given the colour ctσ of cell σ at iteration t, we compute the colour of cell σ at the next iteration ct+1σ by injectively mapping the multi-sets of colours belonging to the adjacent cells of σ using a perfect HASH function: ct+1 3. The algorithm stops when a stable colouring is reached.'), but it is presented in prose rather than a formal pseudocode or algorithm block.
Open Source Code Yes Our code is available at https://github.com/twitter-research/cwn.
Open Datasets Yes CSL Circular Skip Link dataset was first introduced in [57]... SR Similarly to Bodnar et al. [8] and [10]... 3Data available at: http://users.cecs.anu.edu.au/~bdm/data/graphs.html. TUD We test our model on 8 TUDataset benchmarks [56]... ZINC We study the effectiveness of cellular message passing on larger scale molecular benchmarks from the ZINC database [68]... Mol-HIV We additionally test our model on the molecular ogbg-molhiv dataset from the Open Graph Benchmark [40]...
Dataset Splits Yes We follow the same evaluation setting as Dwivedi et al. [24]: 5-fold cross validation procedure and 20 different random weight initialisations. We follow the training and evaluation procedures in [24].
Hardware Specification No The paper states, 'Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Section 5 and Appendix E.' However, the provided text of Section 5 does not contain specific hardware details, and Appendix E is not included.
Software Dependencies No The paper cites PyTorch [59] and PyTorch Geometric [28], indicating their use. However, it does not specify the exact version numbers of these or any other software libraries or dependencies required for replication.
Experiment Setup Yes For simplicity, in all experiments we employ a model which stacks CWN layers with local aggregators as in GIN [74]... See Appendix E for details on feature initialisation, message passing and readout operations, hyperparameters, implementation and benchmark statistics. The training setting and evaluation procedure follow those in Xu et al. [74]. We take the architecture in [27] as reference and replicate the same hyperparameter setting in our model, including the use of only 2 message passing layers.