Discovering Invariant Rationales for Graph Neural Networks
Authors: Yingxin Wu, Xiang Wang, An Zhang, Xiangnan He, Tat-Seng Chua
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on both synthetic and realworld datasets validate the superiority of our DIR in terms of interpretability and generalization ability on graph classification over the leading baselines. |
| Researcher Affiliation | Academia | University of Science and Technology of China National University of Singapore |
| Pseudocode | Yes | Algorithm 1 Pseudocode for DIR in training interpretable Graph Neural Networks (Batch Version) |
| Open Source Code | Yes | Code and datasets are available at https://github.com/Wuyxin/DIR-GNN. |
| Open Datasets | Yes | Code and datasets are available at https://github.com/Wuyxin/DIR-GNN. Spurious-Motif is a synthetic dataset created by following Ying et al. (2019), which involves 18, 000 graphs. MNIST-75sp (Knyazev et al., 2019) converts the MNIST images into 70, 000 superpixel graphs. Graph-SST2 (Yuan et al., 2020; Socher et al., 2013) Each graph is labeled by its sentence sentiment and consists of nodes representing tokens and edges indicating node relations. Molhiv (OGBG-Molhiv) (Hu et al., 2020; 2021; Wu et al., 2017) is a molecular property prediction dataset consisting of molecule graphs, where nodes are atoms, and edges are chemical bonds. |
| Dataset Splits | Yes | Table 3: Statistics of Graph Classification Datasets. Spurious-Motif Train 9,000 Val 3,000 Test 6,000. MNIST-75sp (reduced) Train 20,000 Val 5,000 Test 10,000. Graph-SST2 Train 28,327 Val 3,147 Test 12,305. OGBG-Molhiv Train 32,901 Val 4,113 Test 4,113. |
| Hardware Specification | Yes | All experiments are done on a single Tesla V100 SXM2 GPU (32 GB). |
| Software Dependencies | No | The paper mentions using 'Adam optimizer' but does not provide specific version numbers for any software dependencies or libraries. |
| Experiment Setup | Yes | We set the causal feature ratio and λ as (r = 0.8, λ = 10 4), (r = 0.25, λ = 10 2), (r = 0.6, λ = 102), (r = 0.8, λ = 10 3) for MNIST-75sp, Spurious-Motif, Graph-SST2 and OGBG-Molhiv respectively. For other baselines, we adopt grid search for the best parameters using the validation datasets. The maximum number of epochs is 400 for all datasets. We use Stochastic Gradient Descent (SGD) for the optimization on Graph-SST2 and OGBG-Molhiv and Gradient Descent (GD) for the other two datasets. |