Adaptive Reliable Sample Selection via Prediction Discrepancy for Semi-Supervised Regression

Abstract

Semi-supervised regression has emerged as a critical paradigm to alleviate the dependency on large-scale labeled data, which is often prohibitively expensive or time-consuming to collect in real-world regression tasks. A common practice in semi-supervised regression is to utilize pseudo-labeling, where high-confidence samples are selected based on the model’s estimated uncertainty. However, such uncertainty estimates often do not adequately reflect the true quality of pseudo-labels, leading to the inclusion of low-quality pseudo-labels that degrade model performance. In this paper, we propose a novel sample selection mechanism based on prediction discrepancy between two distinct regressors to filter out low-quality pseudo-labels. To further eliminate the need for manual threshold tuning across various datasets and tasks, we introduce a meta-learning framework to adaptively learn the selection threshold from data. By formulating the threshold acquisition as a bi-level optimization problem, our method adaptively learns the selection threshold using a small set of labeled meta-data. Extensive experiments on three benchmarks demonstrate that our approach effectively identifies reliable unlabeled samples and achieves state-of-the-art performance.

Keywords

Semi-supervised regression / Sample selection / Meta-learning

RIGHTS & PERMISSIONS

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.