Robust Causal Graph Representation Learning against Confounding Effects
Authors: Hang Gao, Jiangmeng Li, Wenwen Qiang, Lingyu Si, Bing Xu, Changwen Zheng, Fuchun Sun
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirically, we conduct extensive experiments on a synthetic dataset and multiple benchmark datasets. Experimental results demonstrate the effectiveness and generalization ability of RCGRL. |
| Researcher Affiliation | Collaboration | Hang Gao1,2*, Jiangmeng Li1,2* , Wenwen Qiang1,2, Lingyu Si1,2, Bing Xu3, Changwen Zheng1, Fuchun Sun4 1Science and Technology on Integrated Information System Laboratory, Institute of Software Chinese Academy of Sciences 2University of Chinese Academy of Sciences 3China Communications Technology Information Group Co., Ltd. 4Tsinghua University {gaohang, jiangmeng2019, qiangwenwen, lingyu, changwen}@iscas.ac.cn, xubing@ccccltd.cn, fcsun@mail.tsinghua.edu.cn |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our codes are available at https://github.com/hang53/RCGRL. |
| Open Datasets | Yes | Datasets. We evaluate our method on both OOD and ID datasets. The OOD datasets include: 1) Spurious-Motif, a synthetic dataset created by (Ying et al. 2019), and we adopt the re-implementation version created by (Wu et al. 2022); 2) Graph-SST2(OOD) is an OOD version of Graph SST2 (Yuan et al. 2020) created by (Wu et al. 2022). The ID datasets include: Graph-SST2 (ID) (Yuan et al. 2020), Graph-Twitter (Yuan et al. 2020), Mol-BBBP, and Mol BACE (Hu et al. 2020). |
| Dataset Splits | Yes | To evaluate the generalization ability of GNN models, we build out-of-domain (OOD) datasets by following the principle of DIR (Wu et al. 2022). Specifically, Spurious-Motif is an artificially generated dataset. Each graph in Spurious-Motif consists of two subgraphs, a ground-truth subgraph, and a confounder subgraph. In the training set, these subgraphs are patched together under certain rules. In contrast, in the validation and test sets, these subgraphs are patched together randomly. |
| Hardware Specification | No | The paper does not provide specific hardware details (exact GPU/CPU models, processor types with speeds, memory amounts, or detailed computer specifications) used for running its experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details (e.g., library or solver names with version numbers) needed to replicate the experiment. |
| Experiment Setup | Yes | Experiment Settings. We compare our method with Empirical Risk Minimization (ERM) and various causality-enhanced methods, including the interpretable baselines, i.e., GAT and Top-k Pool, and the robust learning baselines, i.e., Group DRO, IRM, and DIR. For a fair comparison, we follow the experimental principle of (Wu et al. 2022) and adopt the same training setting for all models, which is described in detail in Appendix. For each task, we report the averaged performance std over ten runs. Some of the results are cited from (Wu et al. 2022). For the values of the hyperparameters, please refer to Appendix. |