ACCon: Angle-Compensated Contrastive Regularizer for Deep Regression

• DOI: 10.1609/aaai.v39i21.34435
• Publication Date: 2025-04-11T13:20:37Z
• Authors: Botao Zhao; Xiaoyang Qu; Zuheng Kang; Junqing Peng; Jing Xiao; Jianzong Wang

In deep regression, capturing the relationship among continuous labels in feature space is a fundamental challenge that has attracted increasing interest. Addressing this issue can prevent models from converging to suboptimal solutions across various regression tasks, leading improved performance, especially for imbalanced and under limited sample sizes. However, existing approaches often rely on order-aware representation learning or distance-based weighting. paper, we hypothesize linear negative correlation between label distances similarities tasks. To implement this, propose an angle-compensated contrastive regularizer which adjusts cosine distance anchor samples within framework. Our method offers plug-and-play compatible solution extends most methods Extensive experiments theoretical analysis demonstrate our proposed not only achieves competitive performance but also excels data efficiency effectiveness datasets.