Mini-Batch Class Composition Bias in Link Prediction
Kieran Maguire, Srinandan Dasmahapatra · 2026
Read Mini-Batch Class Composition Bias in Link Prediction by Kieran Maguire, Srinandan Dasmahapatra (2026) 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-2604-25978.txt.
Description
Maguire and Dasmahapatra show that popular link prediction models can pick up a trivial heuristic that depends on mini-batches. Batch-normalization layers enable it. For a fixed graph, link scores might match node scores when graph traits are shared. That guess fails in general: the trick solves edge sorting from batch stats. Fixing for it lifts fit with node-class traits. The net, it suggests, learns a better graph picture than raw scores imply. Standard link training may thus inflate how general learned pictures are.
Catalog record
| CREATOR | Kieran Maguire, Srinandan Dasmahapatra |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27554 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2604.25978