SpeqNets: Sparsity-aware permutation-equivariant graph networks
Authors: Christopher Morris, Gaurav Rattan, Sandra Kiefer, Siamak Ravanbakhsh
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Here, we aim to empirically investigate the learning performance of the kernel, see Appendix B.1, and neural architectures, see Section 4, based on the (k, s)-LWL, compared with standard kernel and (higher-order) GNN baselines. Concretely, we aim to answer the following questions. Q1 Do the (k, s)-LWL-based algorithms, both kernel and neural architectures, lead to improved classification and regression scores on real-world, graph-level benchmark datasets compared with dense algorithms and standard baselines? Q2 How does the (k, s)-Speq Net architecture compare to standard GNN baselines on node-classification tasks? Q3 To what extent does the (k, s)-LWL reduce computation times compared with architectures induced by the k-WL? Q4 What is the effect of k and s with respect to computation times and predictive performance? The source code of all methods and evaluation procedures is available at https://www.github.com/ chrsmrrs/speqnets. |
| Researcher Affiliation | Academia | Christopher Morris 1 2 3 Gaurav Rattan 1 Sandra Kiefer 4 Siamak Ravanbkash 2 3 1Department of Computer Science, RWTH Aachen University, Aachen, Germany 2Department of Computer Science, Mc Gill University, Montreal, Canada 3Mila, Quebec AI Institute 4Max Planck Institute for Software Systems, Saarland Informatics Campus, Germany. Correspondence to: Christopher Morris <chris@christophermorris.info>. |
| Pseudocode | Yes | Algorithm 1 Generate (k, s)-multisets |
| Open Source Code | Yes | The source code of all methods and evaluation procedures is available at https://www.github.com/ chrsmrrs/speqnets. |
| Open Datasets | Yes | To compare the (k, s)-LWL-based kernels, we used the well-known graph-classification benchmark datasets from (Morris et al., 2020a), see Table 3 for dataset statistics and properties. ... To compare the (k, s)-Speq Net architecture with GNN baselines, we used the ALCHEMY (Chen et al., 2019a) and the QM9 (Ramakrishnan et al., 2014; Wu et al., 2018) graph regression datasets, again see Table 1 for dataset statistics and properties. ... All datasets are publicly available at www. graphlearning.io. ... For both datasets, we uniformly and at random sampled 80% of the graphs for training, and 10% for validation and testing, respectively. |
| Dataset Splits | Yes | Following the evaluation method proposed in (Morris et al., 2020a), the C-parameter was selected from {10 3, 10 2, . . . , 102, 103} using a validation set sampled uniformly at random from the training fold (using 10% of the training fold). Similarly, the numbers of iterations of the (k, s)-LWL, (k, s)-LWL+, 1-WL, WLOA, δ-k-LWL, δ-k LWL+, and k-WL were selected from {0, . . . , 5} using the validation set. ... For both datasets, we uniformly and at random sampled 80% of the graphs for training, and 10% for validation and testing, respectively. ... We used the provided ten training, validation, and test splits for the node-classification datasets. |
| Hardware Specification | Yes | All kernel experiments were conducted on a workstation with 791GB of RAM using a single core. ... All neural experiments were conducted on a workstation with one GPU card with 32GB of GPU memory. |
| Software Dependencies | Yes | All kernels were (re-)implemented in C++11. ... Moreover, we used the GNU C++ Compiler 4.8.5 with the flag -O2. ... We implemented them using PYTORCH GEOMETRIC (Fey & Lenssen, 2019), using a Python-wrapped C++11 preprocessing routine. |
| Experiment Setup | Yes | Following the evaluation method proposed in (Morris et al., 2020a), the C-parameter was selected from {10 3, 10 2, . . . , 102, 103} using a validation set sampled uniformly at random from the training fold (using 10% of the training fold). Similarly, the numbers of iterations of the (k, s)-LWL, (k, s)-LWL+, 1-WL, WLOA, δ-k-LWL, δ-k LWL+, and k-WL were selected from {0, . . . , 5} using the validation set. ... The number of components of the (hidden) node features in {32, 64, 128} and the number of layers in {1, 2, 3, 4, 5} of the GIN and GIN-ε layer were again selected using a validation set sampled uniformly at random from the training fold (using 10% of the training fold). ... We used a 2-layer MLP for the final classification, using a dropout layer with p = 0.5 after the first layer of the MLP. ... we used six layers with 64 (hidden) node features and a set2seq layer (Vinyals et al., 2016) for graph-level pooling, followed by a 2-layer MLP for the joint regression of the twelve targets. |