DKE-RLS: A Manifold Reconstruction Algorithm in Label Spaces with Double Kernel Embedding-Regularized Least Square

Abstract

Dimension reduction is an important preprocess for multi-label classification, such as feature extraction. This paper attempts to explore multi-label learning in the label space. Our approach works using machine learning’s smoothness assumption, where nearby points are more likely to share the same label and the feature manifold and label manifold can share the local topology structure. Thus, here we propose a new multi-label feature-extraction algorithm with a new method for embedding regression, i.e., manifold regularization learning in the subspace formed by multi-labels to reconstruct and use the label manifold. We integrate two least-squares formulas by linear combination, and establish the regression estimation for multi-label manifold learning. To test our approach, we conduct multiple experiments and compare our algorithm against four other multi-label learning algorithms. Results show that our approach significantly improves the performance of label manifold learning.