Class-Imbalanced Graph Learning without Class Rebalancing
Authors: Zhining Liu, Ruizhong Qiu, Zhichen Zeng, Hyunsik Yoo, David Zhou, Zhe Xu, Yada Zhu, Kommy Weldemariam, Jingrui He, Hanghang Tong
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Systematic experiments on real-world imbalanced graph learning tasks show that BAT can deliver up to 46.27% performance gain and up to 72.74% bias reduction over existing techniques. and 4. Experiments We carry out systematic experiments and analysis to validate BAT in the following aspects: (i) Effectiveness in both promoting imbalanced node classification and mitigating the prediction bias between different classes. (ii) Versatility in cooperating with and further boosting various CIGL techniques and GNN backbones. (iii) Robustness to extreme class imbalance. (iv) Efficiency in real-world applications. |
| Researcher Affiliation | Collaboration | 1University of Illinois Urbana-Champaign 2IBM Research 3Amazon Science. |
| Pseudocode | Yes | Algorithm 1 BAT: BAlanced Topological augmentation |
| Open Source Code | Yes | Code, examples, and documentations are available at https://github. com/Zhining Liu1998/BAT. |
| Open Datasets | Yes | We validate BAT on five benchmark datasets for semi-supervised node classification, including the Cora, Cite Seer, Pub Med from Plantoid graphs (Sen et al., 2008), and larger-scale CS, Physics from co-author networks (Shchur et al., 2018) with high-dimensional features. All datasets are publicly available2. https://pytorch-geometric.readthedocs.io/en/latest/modules/datasets.html. |
| Dataset Splits | Yes | We follow the public split (Sen et al., 2008) for semi-supervised node classification where each class has 20 training nodes, then randomly remove minority class training nodes until the given imbalance ratio (IR) is met. and We search for the best architecture for other backbones from #layers l {1, 2, 3} and hidden dimension d {64, 128, 256} based on the average of validation accuracy and F1 score. |
| Hardware Specification | Yes | All the experiments are conducted on a Linux server with Intel Xeon Gold 6240R CPU and NVIDIA Tesla V100 32GB GPU. |
| Software Dependencies | No | The paper mentions using 'pytorch (Paszke et al., 2019)' and 'torch geometric (Fey & Lenssen, 2019)' but does not provide specific version numbers for these software dependencies, which is required for reproducible description. |
| Experiment Setup | Yes | Specifically, we implement all GNNs convolution layer with Re LU activation and dropout (Srivastava et al., 2014) with a dropping rate of 0.5 before the last layer. For GAT, we set the number of attention heads to 4. For APPNP and GPRGNN, we follow the best setting in the original paper and use 2 APPNP/GPR prop convolution layers with 64 hidden units. Note that Graph ENS s official implementation requires modifying the graph convolution for resampling and thus cannot be directly combined with APPNP and GPRGNN. The teleport probability = 0.1 and the number of power iteration steps K = 10. We search for the best architecture for other backbones from #layers l {1, 2, 3} and hidden dimension d {64, 128, 256} based on the average of validation accuracy and F1 score. We train each GNN for 2,000 epochs using Adam optimizer (Kingma & Ba, 2014) with an initial learning rate of 0.01. To achieve better convergence, we follow (Park et al., 2022) to use 5e-4 weight decay and adopt the Reduce LROn Plateau scheduler in Pytorch, which reduces the learning rate by half if the validation loss does not improve for 100 epochs. |