Learning Fuzzy Label-Distribution-Specific Features for Data Processing
Xin Wang, James Dinesh Peter, Adam Słowik, Fan Zhang, Xingsi Xue · IEEE Transactions on Fuzzy Systems · 2024
Due to its superiority in addressing label ambiguity, label distribution learning (LDL) has received wide attention from the community, such as image classification, emotion recognition, and big data processing. To efficiently process the data with label distribution, researchers have proposed to learn label-specific features (LSFs) that are the discriminative features for each class label. Although the LDL literature has seen many algorithms to learn LSFs, most of them ignore the characteristics of label distribution. Label distribution lies in real-value vector space with specific characteristics. In this article, we propose to learn label-distribution-specific features (LDSFs) for processing label distribution data by considering the structures of label distribution. We design a novel LDL method called LDL-LDSF to exploit LDSFs by considering the fuzzy cluster structures of label distribution data. First, LDL-LDSF learns LDSFs for the whole label distribution by jointly learning the label distribution and fuzzy C-means clustering. Second, it learns LDSFs for each label in a similar way. Third, it concatenates the learned LDSFs with the original features to deduce an LDL model. Finally, we conduct extensive experiments to justify that LDL-LDSF statistically outperforms several state-of-the-art LDL methods and validate the advantages of LDSFs for processing label distribution data.