Enhancing Graph Invariant Learning from a Negative Inference Perspective

Authors: Kuo Yang, Zhengyang Zhou, Qihe Huang, Wenjie Du, Limin Li, Wu Jiang, Yang Wang

ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct a comprehensive evaluation of Ne Go on real-world datasets and synthetic datasets across domains. Ne Go outperforms baselines on nearly all datasets, which verify the effectiveness of our framework. ... We conduct extensive experiments on both synthetic and real-world datasets with distribution shifts to evaluate the performance of Ne Go. The results from both visualization and quantitative analysis indicate that our framework successfully achieves accurate prediction in complex environmental scenarios.
Researcher Affiliation Collaboration 1University of Science and Technology of China (USTC), Hefei, China 2Suzhou Institute for Advanced Research, USTC, Suzhou, China 3China Mobile Communications Group Co.,Ltd.
Pseudocode Yes Algorithm 1 The training process of Ne Go
Open Source Code No The paper mentions implementing the model with PyTorch and performing experiments, but does not provide a specific link to the source code, an explicit statement of code release, or indicate that code is included in supplementary materials.
Open Datasets Yes We adopt two synthetic datasets with distribution shift and six real-world scenario shift datasets from both molecular and social science domains. Synthetic datasets include GOOD-Motif (Wu et al., 2022c) and GOOD-CMNIST (Gui et al., 2022). In molecular property prediction fields, we select the scaffold and size splits of GOOD-HIV dataset (Gui et al., 2022; Wu et al., 2018) and the assay and size splits of Drug OOD LBAP-core-ic50 dataset (Ji et al., 2022). We also choose two social sentiment graph datasets with distribution shifts, including GOOD-SST2 and GOODTwitter (Yuan et al., 2022).
Dataset Splits Yes Detailed statistics on the number of graphs in those datasets are provided in Tab. 6. Table 6: Dataset Training ID validation ID test OOD validation OOD test
Hardware Specification Yes We implement our Nego and parts of baselines with Py Torch 1.10.1 on a server with NVIDIA A100-PCIE-40GB.
Software Dependencies Yes We implement our Nego and parts of baselines with Py Torch 1.10.1 on a server with NVIDIA A100-PCIE-40GB.
Experiment Setup Yes During the training stage, we employ the Adam optimizer. We set the maximum number of training epochs to 200. The batch size of training is set as 32 except for GOOD-CMNIST, which uses a batch size of 64. For GOOD-Motif, GOOD-CMNIST and GOODSST2, the learning rate is set to 5 10 4. For GOOD-HIV, GOOD-Twitter, and Drug OOD, we exploit a learning rate of 10 4. Additionally, we utilize a weight decay of 10 4 to help with regularization and prevent overfitting.