Label-guided optimal transport for domain adaptation regression

Domain adaptation regression (DAR) aims to transfer knowledge from a labeled source domain to an unlabeled target domain. Unlike classification tasks that concern a finite set of discrete classes, regression involves a continuous and multidimensional label space, which effectively comprises an infinite number of classes. While this continuity presents significant alignment challenges, it inherently encodes rich semantic information through label relationships. Regrettably, existing methods often fail to exploit this inherent advantage, focusing solely on feature alignment while neglecting rich structural information embedded within continuous label relationships. To overcome these limitations, we propose a Label-Guided Optimal Transport (LGOT) framework. Our method leverages the geometric properties of optimal transport to align features and integrate label semantics directly into the feature alignment process. The core of LGOT is a Similarity-Driven Adaptive Threshold mechanism, which dynamically computes the thresholds between sample pairs based on their label similarity. By assigning lower costs to more similar pairs, our model encourages their features to be pulled closer, effectively transferring the continuous structure of the label space into meaningful constraints for the learned feature space. Extensive experiments on three standard regression benchmarks demonstrate the effectiveness of the proposed method. © 2026 Elsevier Ltd.