IceBerg: Debiased Self-Training for Class-Imbalanced Node Classification
Zhixun Li, Dingshuo Chen, Tong Zhao, Daixin Wang, Hongrui Liu, Zhiqiang Zhang, Jun Zhou †, Jeffrey Xu Yu · 2025
Read IceBerg: Debiased Self-Training for Class-Imbalanced Node Classification by Zhixun Li, Dingshuo Chen, Tong Zhao, Daixin Wang, Hongrui Liu, Zhiqiang Zhang, Jun Zhou †, Jeffrey Xu Yu (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-06280.txt.
Description
Li and colleagues wrote in twenty twenty-five about graph neural networks that must learn from skewed class labels. In many graphs, a few classes hold most labeled nodes. Unlabeled nodes are plentiful but often ignored by past fixes. Their framework, IceBerg, trains the model partly on its own predictions for unlabeled nodes. Two problems dog that approach: bias toward big classes and shifted label patterns. A double balancing module tackles both with only a few lines of code.
They also split message passing from feature change to spread weak signals farther. Tests on standard benchmarks beat prior imbalance methods clearly. The framework also led the field when labels were extremely scarce. The authors argue unlabeled data are the bulk of the iceberg. Small, careful changes can unlock large gains, they conclude.
Catalog record
| CREATOR | Zhixun Li, Dingshuo Chen, Tong Zhao, Daixin Wang, Hongrui Liu, Zhiqiang Zhang, Jun Zhou †, Jeffrey Xu Yu |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 52 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-06280