Contrastive Order Learning for Ordinal Regression

Chaewon Lee, BeomJun Shim, Kwang Pyo Choi, Chang-Su Kim

Keywords: contrastive learning, order learning, ordinal regression

Abstract: We propose a novel contrastive learning algorithm for ordinal regression, called contrastive order learning (ConOrd), which combines the strengths of order learning and contrastive learning effectively. While contrastive learning excels at leveraging all samples in a batch, it often overlooks the inherent order among rank labels. Conversely, order learning superbly captures label ordinality but typically relies on local, margin-based comparisons, limiting its global consistency and representation power. ConOrd addresses these limitations by introducing a contrastive order loss, which adopts soft affinity and disparity weights based on rank differences, enabling fine-grained modeling of ordinal relationships across all sample pairs in a batch. Moreover, to enhance the embedding space further, we incorporate a center loss with learnable reference points, which guide compact clustering and ordinal alignment. Extensive experiments on various ordinal regression tasks, including facial age estimation, blind image quality assessment, and blind video quality assessment, demonstrate that the proposed ConOrd consistently provides state-of-the-art results and generalizes well to diverse ordinal regression scenarios.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 12479