Learning sample representativeness for class-imbalanced multi-label classification

• Published in: Pattern analysis and applications : PAA Vol. 27; no. 2
• Main Authors: Zhang, Yu, Cao, Sichen, Mi, Siya, Bian, Yali
• Format: Journal Article
• Language: English
• Published: London Springer London 01.06.2024; Springer Nature B.V
• Subjects: Classification; Computer Science; Datasets; Image classification; Labels; Original Article; Parameter modification; Pattern Recognition; Production methods; Multi-label classification; Class imbalance; Re-weighting representative samples
• ISSN: 1433-7541; 1433-755X
• DOI: 10.1007/s10044-024-01209-8

Class imbalance is a common problem that often occurs in multi-label image classification. In multi-label datasets, the co-occurrence of labels presents a unique set of difficulties, making it hard for traditional methods to produce satisfactory results, particularly on tail classes. Based on previous research and our investigation, we have found that the number of labels presents in a given sample can influence classification results. Nevertheless, it is worth noting that certain samples within the tail classes exhibit resistance to this influence, which is a critical aspect in the context of class-imbalanced multi-label classification. In this paper, we term these samples as representative samples. Highlighting representative samples during training can effectively address the above issues. Specifically, we propose a new method to learn sample representativeness, which is named Representativeness-Emphasizing Loss (REL). First, we use a new re-weighting form to rebalance the weights based on sample representativeness. Then, a modified focal loss dynamically assigns tailored parameters for each class in each sample to further emphasize the sample representativeness. Extensive experiments on two class-imbalanced datasets show that models trained with this new loss function achieve comparable performance to existing methods.