CASSOR: Class-Aware Sample Selection for Ordinal Regression with Noisy Labels

Yue Yuan, Sheng Wan, Chuang Zhang, Chen Gong

Abstract: Ordinal regression aims at solving the classification problem, where the categories are related in a natural order. Due to the difficulty in distinguishing between highly relevant categories, label noise is frequently present in ordinal data. Moreover, the varying degrees of relevance between categories can lead to an inconsistent distribution of misclassification loss across categories, posing a challenge to select clean data consistently from all categories for training. To overcome this limitation, we develop a sample selection method termed ‘Class-Aware Sample Selection for Ordinal Regression’ (CASSOR). To be concrete, we devise a class-specific sample selection strategy in order to adaptively acquire sufficient clean examples for robust model training. Moreover, a label-ranking regularizer is designed to help guide the sample selection process via exploring the ordinal relationship between different examples. As a result, our proposed CASSOR is endowed with strong discrimination abilities on ordinal data. Intensive experiments have been performed on multiple real-world ordinal regression datasets, which firmly demonstrates the effectiveness of our method.