We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Machine Learning

Title: Simple data balancing achieves competitive worst-group-accuracy

Abstract: We study the problem of learning classifiers that perform well across (known or unknown) groups of data. After observing that common worst-group-accuracy datasets suffer from substantial imbalances, we set out to compare state-of-the-art methods to simple balancing of classes and groups by either subsampling or reweighting data. Our results show that these data balancing baselines achieve state-of-the-art-accuracy, while being faster to train and requiring no additional hyper-parameters. In addition, we highlight that access to group information is most critical for model selection purposes, and not so much during training. All in all, our findings beg closer examination of benchmarks and methods for research in worst-group-accuracy optimization.
Comments: Accepted at CLeaR (Causal Learning and Reasoning) 2022
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Cryptography and Security (cs.CR)
Cite as: arXiv:2110.14503 [cs.LG]
  (or arXiv:2110.14503v2 [cs.LG] for this version)

Submission history

From: Badr Youbi-Idrissi [view email]
[v1] Wed, 27 Oct 2021 15:15:11 GMT (3980kb,D)
[v2] Fri, 18 Feb 2022 17:07:14 GMT (4289kb,D)

Link back to: arXiv, form interface, contact.