A Semi-Supervised Ordinal Classification Method Based on Ranking Consistency Regularization

We propose a semi-supervised ordinal classification method based on ranking consistency regularization, addressing limitations in capturing ordinal relationships and mitigating semantic feature disruption from data augmentation. The method leverages Ordinal Unimodal Cross-Entropy (OUCE) loss to enforce unimodality in predicted probability distributions, ensuring robust ordinal constraints. An Auxiliary Ranking Classifier (ARC) is designed to explicitly model ordinal relationships between sample pairs. The ARC is combined with ranking consistency regularization across strongly and weakly augmented views to enhance prediction stability under perturbations. Additionally, the class-adaptive thresholding strategy dynamically adjusts class-specific thresholds via exponential moving averages (EMA) to effectively mitigate confirmation bias. Experiments on the Adience and OIA-DDR datasets demonstrate competitive performance.