New Regularization Method Improves Ordinal Regression Performance

This research addresses a challenge in ordinal regression, a type of classification where target variables have a natural ordered relationship. Many real-world ordinal datasets exhibit unimodal conditional probability distributions (CPDs), meaning there's a single peak in the probability of a target variable given an explanatory variable. Previous unimodality-promoting regularized learning (UPRL) methods aimed to leverage this property to reduce prediction variance. However, the study found that existing UPRL methods inadvertently promote CPDs to become not only more unimodal but also smoother or less confident, introducing an unintended scale-related bias. To counteract this, the researchers developed a novel UPRL method that more precisely enforces unimodality without inducing this scale bias. Experimental comparisons confirm that this new method effectively improves prediction performance, particularly when dealing with smaller training datasets. The analysis explains how the presence or absence of the scale-related bias accounts for the observed performance differences, demonstrating that a purer unimodality promotion indeed contributes to better results in ordinal regression.