Learning from interval targets

Rattana Pukdee, Ziqi Ke, Chirag Gupta

Keywords: Weak supervision, Partial-label learning, Learning from side information, Learning theory

Abstract: We consider regression problems where the exact real-valued targets are not directly available; instead, supervision is provided in the form of intervals around the targets—that is, only lower and upper bounds are known. Such a "learning from interval targets" setup arises in domains where labeling costs are high or there is inherent uncertainty in the target values. In these settings, traditional regression loss functions, which require exact target values, cannot be directly applied. To address this challenge, we propose two approaches: (i) modifying the regression loss function to be compatible with interval ground truths, and (ii) formulating a min-max problem where we minimize the typical regression loss with respect to the "worst-case" label within the interval. We provide theoretical guarantees for our methods, analyze their computational efficiency, and evaluate their practical performance on real-world datasets.

Primary Area: learning theory

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 3736