Ordinal Regression by Combining Finely Labeled Data with Coarser-Labeled Data

Abstract

Ordinal regression is a classification problem with an order relationship between labels. In classification, including ordinal regression, finely labeled and coarser-labeled data are sometimes obtained because of different types of collection or time periods. Ideally, all obtained data can be used for training models, but it is impossible to train models directly using data with differently labeled granularity. We converted two different-granularity labels to soft labels based on a specific probability distribution. This enabled coarser-labeled data to be added to finely labeled data for training. We then performed experiments on two benchmark datasets to evaluate our proposed method. We found the method improves the performance of LightGBM in the ordinal regression problem by up to about 0.02 in accuracy and about 0.04 in mean absolute error. Additionally, a smaller probability distribution variance tends to improve accuracy, while a larger one tends to improve the mean absolute error.