Towards Label Position Bias in Graph Neural Networks
Authors: Haoyu Han, Xiaorui Liu, Feng Shi, MohamadAli Torkamani, Charu Aggarwal, Jiliang Tang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our proposed method not only outperforms backbone methods but also significantly mitigates the issue of label position bias in GNNs. In this section, we conduct comprehensive experiments to verify the effectiveness of the proposed LPSL. |
| Researcher Affiliation | Collaboration | Haoyu Han1, Xiaorui Liu2, Feng Shi3, Mohamad Ali Torkamani4 , Charu C. Aggarwal5, Jiliang Tang1 1Michigan State University 2North Carolina State University 3Tiger Graph 4 Amazon 5IBM T.J. Watson Research Center {hanhaoy1,tangjili}@msu.edu, xliu96@ncsu.edu bill.shi@tigergraph.com, alitor@amazon.com, charu@us.ibm.com |
| Pseudocode | Yes | Algorithm 1 Algorithm of LPSL |
| Open Source Code | Yes | Our code is available at: https://github.com/haoyuhan1/LPSL. |
| Open Datasets | Yes | We conduct experiments on 8 real-world graph datasets for the semi-supervised node classification task, including three citation datasets, i.e., Cora, Citeseer, and Pubmed [27], two coauthorship datasets, i.e., Coauthor CS and Coauthor Physics, two co-purchase datasets, i.e., Amazon Computers and Amazon Photo [28], and one OGB dataset, i.e., ogbn-arxiv [29]. |
| Dataset Splits | Yes | For label rates 5, 10, and 20, we use 500 nodes for validation and 1000 nodes for testing. For label rates of 60% labeled node per class, we use half of the rest nodes for validation and the remaining half for the test. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory, or cloud instance types) used to conduct the experiments. |
| Software Dependencies | No | The paper mentions software like "Adam optimizer" and cites "PyTorch Geometric" but does not provide specific version numbers for these or other software dependencies. |
| Experiment Setup | Yes | For the proposed LPSLGCN, we set the λ in range [1,8]. For LPSLAPPNP, we set the λ in the range [8, 15]. For both methods, c is set in the range [0.5, 1.5]. We fix the learning rate 0.01, dropout 0.5 or 0.8, hidden dimension size 64, and weight decay 0.0005, except for the ogbn-arxiv dataset. More details about the hyperparameters setting for all methods can be found in Appendix D. |