Contrastive Order Learning Improves Ordinal Regression Performance

A new framework called Contrastive Order Learning (ConOrd) has been developed to significantly improve ordinal regression tasks. Ordinal regression involves predicting labels that have a natural order, such as age groups or quality ratings. While contrastive learning is powerful for leveraging relationships between data samples, it typically overlooks the specific sequential nature of ordered labels. ConOrd addresses this by combining the strengths of contrastive learning with explicit order modeling. It introduces a unique contrastive order loss function that uses "soft affinity" and "disparity weights." These weights are dynamically adjusted based on the differences in rank between sample pairs within a batch, allowing the model to capture fine-grained ordinal relationships across all samples. Extensive testing across various applications, including estimating facial age, assessing blind image quality, and evaluating blind video quality, demonstrated that ConOrd consistently outperforms existing methods. This indicates its strong generalization capabilities and potential for broad application in scenarios requiring precise ordinal predictions.