Nonlinear Characteristic-Driven Partial Multi-label Learning

Partial Multi-label Learning (PML) addresses challenges where each instance is associated with multiple candidate labels, which include both relevant and irrelevant labels. Traditional label disambiguation strategies often overlook the significance of data lying on nonlinear subspace. Assuming that data points closely adhere to multiple linear subspaces is restrictive and may fail in many applications. Linear subspace clustering algorithms frequently struggle with data residing on multiple nonlinear manifolds, as they inherently focus solely on the global linear associations between data points. To address this gap, we propose a novel approach called Nonlinear Characteristic-Driven Partial Multi-Label Learning (NLPML). This method introduces a kernel low-rank representation technique to learn the instance relationship matrix within a nonlinear subspace, effectively identifying instances with complex nonlinear structures. Simultaneously, we introduce the Jaccard distance to quantify label relevance by constructing a label affinity matrix. By jointly optimizing instance-level and label-level affinity matrices, NLPML leverages complementary information for effective label denoising. Experimental results demonstrate that NLPML significantly outperforms existing methods.