Histogram-Guided Source-Free Domain Adaptive Regression

ICLR 2026 Conference Submission12513 Authors

Keywords: Domain adaptation, unsupervised learning, regression model

Abstract: Source-Free Domain Adaptation (SFDA) enables model adaptation under distribution shifts without access to source data, making it an appealing solution for privacy-sensitive applications. Despite being a fundamental problem in machine learning, regression remains largely underexplored in SFDA, where most existing work has focused predominantly on classification tasks. To bridge this gap, we propose a novel algorithm that leverages sample-wise, histogram-informed supervisory signals to refine pseudo-labels under an uncertainty-aware paradigm. This design simultaneously achieves pseudo-label refinement and uncertainty modeling, two key components that are critical for effective adaptation in classification but remain largely absent in regression. We further theoretically show that the resulting histograms exhibit robustness to potential perturbations, supporting reliable SFDA for regression. Empirical results across multiple benchmarks confirm the effectiveness of our method and reveal that histogram-guided learning promotes more compact and structured feature representations, mitigating the inherent challenges of adapting regression models under distribution shift.

Primary Area: transfer learning, meta learning, and lifelong learning

Submission Number: 12513