Multi-task ordinal regression with labeled and unlabeled data

Ordinal regression (OR) aims to construct the classifier from data with ordered class labels. At present, most of the OR methods consider the OR problem as a single learning task and assume that all samples in the learning task are labeled. Nevertheless, in practice, labeling a large number of samples may be costly. If there are not enough labeled samples available to train the classifier, their classification accuracy may be restricted. To deal with this problem, in this paper, we propose a novel multi-task semi-supervised ordinal regression (MTSSOR) method. Our method is able to incorporate the additional information from the related tasks and unlabeled samples into improving the performance of OR classifiers, when the labeled samples are insufficient. To the best of our knowledge, this is the first attempt on multi-task semi-supervised OR. In the experiments, we compare MTSSOR with the single-task supervised OR methods, single-task semi-supervised OR methods and multi-task supervised OR method on real-world multi-task OR datasets. The mean-zero error (MZE) and mean-absolute error (MAE) are used as evaluation metrics. The experimental results show that compared with these OR methods, MTSSOR can achieve a minimum of 0.015 and up to 0.152 improvements in term of MZE, and a minimum of 0.02 and up to 0.272 improvements in term of MAE. •The multi-task ordinal regression problem is addressed.•The unlabeled samples are incorporated into the learning process via manifold regularization.•The knowledge is shared among different tasks by setting a shared norm vector.