Rethinking the Setting of Semi-supervised Learning on Graphs
Authors: Ziang Li, Ming Ding, Weikai Li, Zihan Wang, Ziyu Zeng, Yukuo Cen, Jie Tang
IJCAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments suggest that IGB is a more stable benchmark than previous datasets for semi-supervised learning on graphs. Our code and data are released at https://github.com/THUDM/IGB/. |
| Researcher Affiliation | Academia | Department of Computer Science and Technology, Tsinghua University {li-za19, dm18, liwk19, zhwang19, zengzy19, cyk20}@mails.tsinghua.edu.cn jietang@tsinghua.edu.cn |
| Pseudocode | Yes | The full pipeline of Valid Util is defined as follows and in Algorithm 1 |
| Open Source Code | Yes | Our code and data are released at https://github.com/THUDM/IGB/. |
| Open Datasets | Yes | IGB consists of four datasets: AMiner [Tang et al., 2008], Facebook [Rozemberczki et al., 2019], NELL [Yang et al., 2016], and Flickr [Zeng et al., 2019]. |
| Dataset Splits | Yes | Cora 2,708 5,429 140 / 500 / 1,000 (from Table 1) and Divide the labeled set into training and validation sets.3 The best ratio may differ from model to model. In practice, we find that 1:1 is an appropriate ratio for most models. (Section 3.2) |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments. |
| Software Dependencies | No | The paper states “The experiments can be reproduced using the Cog DL package [Cen et al., 2021]” and mentions PyG, but does not provide specific version numbers for any software dependencies. |
| Experiment Setup | Yes | The search scope of the hyper-parameters includes learning rate, hidden size, early stopping iteration, number of layers, the dropout rate, the diffusion radius of APPNP and GDC, the sparsification threshold of GDC, etc. We use grid search to find the best hyper-parameters for each model. Details about the search scopes are shown in the released codes. We set the dropout rate as 0 when searching for the best pseudo-labels. |