Graph Sparsification via Mixture of Graphs
Authors: Guibin Zhang, Xiangguo SUN, Yanwei Yue, Chonghe Jiang, Kun Wang, Tianlong Chen, Shirui Pan
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on four large-scale OGB datasets and two superpixel datasets, equipped with five GNN backbones, demonstrate that Mo G (I) identifies subgraphs at higher sparsity levels (8.67% 50.85%), with performance equal to or better than the dense graph, (II) achieves 1.47 2.62 speedup in GNN inference with negligible performance drop, and (III) boosts top-student GNN performance (1.02% on Rev GNN+OGBNPROTEINS and 1.74% on Deeper GCN+OGBG-PPA). |
| Researcher Affiliation | Academia | 1Tongji University 2CUHK 3NTU 4UNC-Chapel Hill 5Griffith University |
| Pseudocode | Yes | Algorithm 1: Algorithm workflow of Mo G |
| Open Source Code | Yes | The source code is available at https://github.com/yanweiyue/Mo G. |
| Open Datasets | Yes | We opt for four large-scale OGB benchmarks (Hu et al., 2020), including OGBN-ARXIV, OGBN-PROTEINS and OGBN-PRODUCTS for node classification, and OGBG-PPA for graph classification. The dataset splits are given by (Hu et al., 2020). Additionally, we choose two superpixel datasets, MNIST and CIFAR-10 (Knyazev et al., 2019). |
| Dataset Splits | Yes | For node-level tasks, the data splits for OGBN-ARXIV, OGBN-PROTEINS, and OGBN-PRODUCTS were provided by the benchmark (Hu et al., 2020). Specifically, for OGBN-ARXIV, we train on papers published until 2017, validate on papers from 2018 and test on those published since 2019. For OGBN-PROTEINS, protein nodes were segregated into training, validation, and test sets based on their species of origin. For OGBN-PRODUCTS, we sort the products according to their sales ranking and use the top 8% for training, next top 2% for validation, and the rest for testing. For graph-level tasks, we follow (Hu et al., 2020) for OGBG-PPA. Concretely, we adopt the species split, where the neighborhood graphs in the validation and test sets are extracted from protein association networks of species not encountered during training but belonging to one of the 37 taxonomic groups. This split stress-tests the model s capacity to extract graph features crucial for predicting taxonomic groups, enhancing biological understanding of protein associations. For MNIST and CIFAR-10, consistent with (Dwivedi et al., 2020), we split them to 55000 train/5000 validation/10000 test for MNIST, and 45000 train/5000 validation/10000 test for CIFAR10, respectively. We report the test accuracy at the epoch with the best validation accuracy. |
| Hardware Specification | Yes | All the experiments are conducted on NVIDIA Tesla V100 (32GB GPU), using Py Torch and Py Torch Geometric framework. |
| Software Dependencies | No | All the experiments are conducted on NVIDIA Tesla V100 (32GB GPU), using Py Torch and Py Torch Geometric framework. While software names are mentioned, specific version numbers for PyTorch and PyTorch Geometric are not provided, which is required for reproducibility. |
| Experiment Setup | Yes | For node classification backbones, we utilize a 3-layer Graph SAGE with hidden_dim {128, 256}. As for Deeper GCN, we set layer_num = 28, block = res+, hidden_dim = 64. The other configurations are the same as in https://github.com/lightaime/deep_gcns_torch/tree/master/ examples/ogb/ogbn_proteins. For graph classification backbones, we leverage a 4-layer PNA with hidden_dim = 300. Rest configurations are the same as in https://github.com/ lukecavabarrett/pna. Mo G parameters We adopt the m = 4 sparsity criteria outlined in Section 3.3, assigning n = 3 different sparsity levels {s1, s2, s3} to each criterion, resulting in a total of K = m n = 12 experts. We select k = 2 sparsifier experts for each node, and set the loss scaling factor λ = 1e 2 across all datasets and backbones. |