Progressive label enhancement

Label Distribution Learning (LDL) leverages label distribution (LD) to represent instances, which helps solve label ambiguity. However, obtaining LD can be extremely challenging in many real-world scenarios. Label Enhancement (LE) has emerged as a solution to enhance logical labels to LD since logical labels are highly available. In this paper, we explore the application of dimension reduction techniques to enhance LE. We present a learning framework known as Progressive Label Enhancement (PLE). PLE progressively conducts dependency-maximization-oriented dimension reduction and LE. First, PLE generates LD by leveraging the manifold structure within the feature space induced by dependency-maximization-driven dimension reduction. Second, PLE optimizes the projection matrix for dependency maximization based on the obtained LD. Finally, extensive experiments conducted on 15 real-world datasets consistently demonstrate that PLE outperforms the other six comparative approaches.