Multi-Label Learning based on Label-Specific Features and Single-Annulus Clustering
Yan Liu, Jingjing Song, Taihua Xu, Xuan Yan, Eric C.C. Tsang · 2022
The label-specific features learning that extracts specific features from different labels for classification has attracted wide attention in recent years. The strategy of label-specific features is to exploit customized features in place of the original feature representation in their discrimination processes. However, there is a massive bias between the number of positive and negative instances in most multi-label data. In addition, the cross-distribution of instances and unbalanced label density further increase the difficulty of multi-label classification simultaneously. In this paper, we present a novel strategy called single-annulus clustering to form hierarchical label-specific features that prevent instances of unstable aggregation in the case of significant bias and improve classification efficiency. The principal objective of this strategy is to bisect instances for each label into annuluses with different sizes and then extract underlying features within each layer of annuluses. Experiments on benchmark multi-label data sets and contrastive research against other well-known algorithms show the effectiveness of the proposed strategy.