Debiasing Graph Neural Networks via Learning Disentangled Causal Substructure
Authors: Shaohua Fan, Xiao Wang, Yanhu Mo, Chuan Shi, Jian Tang
NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results well demonstrate that our approach achieves superior generalization performance over existing baselines. Furthermore, owing to the learned edge mask, the proposed model has appealing interpretability and transferability. |
| Researcher Affiliation | Academia | 1Beijing University of Posts and Telecommunications, China 2 Mila Québec AI Institute, Canada 3 HEC Montréal, Canada 4 CIFAR AI Research Chair |
| Pseudocode | No | The paper does not include a clearly labeled pseudocode or algorithm block. While it states “Our approach is summarized in App. B,” Appendix B provides a textual summary rather than a formal pseudocode listing. |
| Open Source Code | Yes | Code and data are available at: https://github.com/googlebaba/Dis C. |
| Open Datasets | Yes | Code and data are available at: https://github.com/googlebaba/Dis C. We construct three new graph datasets, which have controllable bias degrees and are easier to visualize and explain. Following CMNIST-75sp introduced in Sec. 3.1, we use the similar way to construct CFashion-75sp and CKuzushiji-75sp datasets based on the Fashion-MNIST [40] and Kuzushiji-MNIST [4] datasets. |
| Dataset Splits | Yes | The training, validation and test splits are 70%, 10% and 20% respectively. For the testing set, we construct both biased testing set and unbiased testing set. The biased testing set has the same bias degree with training set, aiming to measure the extent of models relying on bias. The unbiased testing set, where the digit labels uncorrelate with the background colors, aims to test whether the model could utilize the inherent digit signals for prediction. |
| Hardware Specification | No | The paper states: “The computation resource of this project is supported by Compute Canada.” While Compute Canada is a provider of computational resources, this statement does not specify the exact GPU models, CPU types, or other detailed hardware specifications used for the experiments. |
| Software Dependencies | No | The paper mentions using and comparing against various GNN models (GCN, GIN, GCNII, Diff Pool, Factor GCN, DIR, Stable GNN, LDD) but does not provide specific version numbers for any software dependencies, such as Python, PyTorch, TensorFlow, or other libraries. |
| Experiment Setup | Yes | To keep fair comparison, our model uses the same GNN architecture and hyperparameters with the corresponding base model. All the experiments are run 4 times with different random seeds and we report the accuracy and the standard error. Hyperparameter experiments Fig. 5 is the hyperparameter experiments of the degree of amplifying bias q in GCE loss and the importance of generation component λG. For q, we fix λG = 10 and vary q from {0.1, 0.3, 0.5, 0.7, 0.9}. For λG, we fix q = 0.7 and vary λG from {1, 5, 10, 15}. More details are in App. C.2 (which specifies learning rates 0.0005/0.001, hidden dimensions 128, batch size 64, dropout 0.5, epochs 100). |