Improving the classification of extreme classes by means of loss regularisation and generalised beta distributions

TL;DR

This work tackles ordinal classification with a focus on reliably predicting the extreme classes. It extends unimodal regularisation by introducing a generalised beta distribution GB$(\alpha,u,v)$ for soft-label encoding, using $\alpha=2$ for the extremes to tighten mass near 0 and 1 while keeping intermediates centered with $\alpha=1$. A tractable parameter-estimation scheme fixes $\alpha$ and derives $u$ and $v$ through mean/variance constraints, augmented by class-specific strategies for intermediate and extreme classes. Evaluated on six image-based ordinal benchmarks with a ResNet18 backbone, the proposed GB regularisation improves extreme-class performance (GMSEC) and remains competitive on standard metrics, supported by statistical tests that confirm significance. The approach offers a practical, principled way to prioritize accuracy on the most critical classes in ordinal tasks.

Abstract

An ordinal classification problem is one in which the target variable takes values on an ordinal scale. Nowadays, there are many of these problems associated with real-world tasks where it is crucial to accurately classify the extreme classes of the ordinal structure. In this work, we propose a unimodal regularisation approach that can be applied to any loss function to improve the classification performance of the first and last classes while maintaining good performance for the remainder. The proposed methodology is tested on six datasets with different numbers of classes, and compared with other unimodal regularisation methods in the literature. In addition, performance in the extreme classes is compared using a new metric that takes into account their sensitivities. Experimental results and statistical analysis show that the proposed methodology obtains a superior average performance considering different metrics. The results for the proposed metric show that the generalised beta distribution generally improves classification performance in the extreme classes. At the same time, the other five nominal and ordinal metrics considered show that the overall performance is aligned with the performance of previous alternatives.

Figures (4)

• Figure 1: Probability density functions of the GB distributions for a problem with five classes. In the extreme classes, the standard beta ($\alpha=1$) is compared to the proposed alternative ($\alpha=2$).
• Figure 2: One image of each category ($0$ to $4$ from left to right) obtained from the training set of the diabetic retinopathy dataset.
• Figure 3: Images obtained from the AVA dataset.
• Figure 4: Images extracted from the different categories of the WIKI dataset.