Aiyappa et al., "Implicit degree bias in the link prediction task" (2025)
2025-10-06 → 2026-08-14
Rachith Aiyappa, Xin Wang, Munjung Kim, Ozgur Can Seckin, Yong-Yeol Ahn, and Sadamori Kojaku, ICML (2025)
ICML | arXiv | Code | Data
@inproceedings{aiyappa2025implicit,
author = {Rachith Aiyappa and Xin Wang and Munjung Kim and Ozgur Can Seckin and Yong-Yeol Ahn and Sadamori Kojaku},
title = {Implicit degree bias in the link prediction task},
booktitle = {Proceedings of the 42nd International Conference on Machine Learning (ICML)},
volume = {267},
pages = {874--908},
series = {Proceedings of Machine Learning Research},
publisher = {PMLR},
address = {Vancouver, Canada},
month = {7},
archivePrefix = {arXiv},
eprint = {2405.14985},
primaryClass = {cs.SI},
year = {2025},
}
Standard Link prediction benchmarks contain an inherent bias favoring high-degree nodes, allowing simple degree-based methods to achieve near-optimal results. We introduce a degree-corrected variant of the task that provides fairer evaluation and better alignment with real-world recommendation performance. The revised benchmark improves model training by reducing overfitting to node degrees and enabling better learning of meaningful graph structures.