Label-dependent feature exploration for label distribution learning

Abstract

Label distribution learning (LDL) explicitly models label ambiguity by assigning a real-valued vector with label description degrees to each sample. Most LDL methods only build models on the same feature (sub)space shared by all labels. However, they ignore that each label has its own specific features, and there are some common features among labels. In this paper, we propose a novel LDL (LDL-LDF) algorithm that aims to exploit both label-dependent and common features. First, label-dependent feature reconstruction utilizes thresholding for relevant sample subset identification, density peaks clustering for representative sample selection, and Euclidean distance for feature value calculation. Second, common feature reconstruction follows a similar approach, however, on the whole dataset. Finally, the prediction neural network is composed of several components that serve each label with label-dependent features, one component that serves all labels with common features, and the fusion component. The effectiveness and competitiveness of our algorithm are verified through various experiments comparing seven algorithms on fourteen real-world datasets.