Ranking Enhanced Supervised Contrastive Learning for Regression

Supervised contrastive learning has shown promising results in image classification tasks where the representations are pulled together if they share same labels or otherwise pushed apart. Such dispersion process in the representation space benefits the downstream classification tasks. However, when applied to regression tasks directly, such dispersion lacks guidance of the relationship among target labels (i.e. the label distances), which leads to the disalignment between representation distances and label distances. Achieving such alignment without compromising the dispersion of learned representations is challenging. In this paper, we propose a Ranking Enhanced Supervised Contrastive Loss (RESupCon) to empower the representation dispersion process with ranking alignment between representation distances and label distances in a controlled fashion. We demonstrate the effectiveness of our method in image regression tasks on four real-world datasets with various interests, including meteorological, medical and human facial data. Experimental results of our method show that representations with better ranking are learned and improvements are made over other baselines in terms of RMSE on all four datasets.