Weisfeiler-Leman at the margin: When more expressivity matters

Authors: Billy Joe Franks, Christopher Morris, Ameya Velingker, Floris Geerts

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical study confirms the validity of our theoretical findings. and 5. Experimental evaluation and Results and discussion In the following, we answer questions Q1 to Q4.
Researcher Affiliation Collaboration 1Department of Computer Science, University of Kaiserslautern-Landau 2Department of Computer Science, RWTH Aachen University 3Google Research 4Department of Computer Science, University of Antwerp.
Pseudocode No The paper describes algorithms in prose and mathematical notation but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks.
Open Source Code Yes The source code of all methods and evaluation procedures is available at https://www.github.com/chrsmrrs/wl_vc_expressivity.
Open Datasets Yes We used the well-known graph classification benchmark datasets from Morris et al. (2020a); see Table 2 for dataset statistics and properties.2All datasets are publicly available at www.graphlearning.io.
Dataset Splits Yes In both cases, we used 10-fold cross-validation. We repeated each 10-fold cross-validation ten times with different random folds and report average training and testing accuracies and standard deviations. and using a validation set sampled uniformly at random from the training fold (using 10 % of the training fold).
Hardware Specification Yes All kernel experiments were conducted on a workstation with 512 GB of RAM using a single CPU core. and All MPNN experiments were conducted on a workstation with 512 GB of RAM using a single core and one NVIDIA Tesla A100s with 80 GB of GPU memory.
Software Dependencies No We implemented the (normalized) 1-WL, 1-WLOA, 1-WLF, and the 1-WLOAF in Python. and computed the classification accuracies using the C-SVM implementation of LIBSVM (Chang & Lin, 2011). and computed the classification accuracies using the linear SVM implementation of LIBLINEAR (Fan et al., 2008).
Experiment Setup Yes For the experiments on the TUDATSETS, following the evaluation method proposed in Morris et al. (2020a), the C-parameter and numbers of iterations were selected from {10 3, 10 2, . . . , 102, 103} and {1, . . . , 5}, respectively, using a validation set sampled uniformly at random from the training fold (using 10 % of the training fold). and We used an initial learning rate of 0.01 across all experiments with an exponential learning rate decay with patience of 5, a batch size of 128, and set the maximum number of epochs to 200. and We used mean pooling and a two-layer MLP using a dropout of 0.5 after the first layer for all experiments for the final classification.