Dual Subspace Representation and Second-Order Affinity Graph Collaboration for Partial Multi-Label Feature Selection
Year: 2026
Type: article
Abstract: Feature selection is an effective technique for solving curse of dimensionality in high-dimensional data. Lots of models emerged for tackling multi-label classification tasks, but most of them perform inferiorly in the scenarios of partial multi-label learning (PML), where data annotation contains noisy labels. Although the commonly used matrix decomposition technique can suppress noise in observed labels to some extent, existing partial multi-label feature selection (PMFS) methods usually exhibit insufficient exploration of interface between feature space and label space. To address this issue, a dual subspace representation and second-order affinity graph collaboration (DSRSAG) method is proposed. In the model, the observed multiple labels are decomposed into low-rank ground-truth labels and sparse noise via the robust principal component analysis (RPCA). The dual subspace representations between instances and ground-truth labels are modeled to uncover consistent structures of instances and labels. Meanwhile, an adaptive second-order affinity graph is introduced to address the absence of first-order correlations of instances and mitigate the negative impact of noise on the model. Additionally, a novel non-convex function is constructed to better reveal the low-rank structure of ground-truth label matrix. To exclude the influence of non-essential features, the$\ell _{2,1}$-norm is imposed on feature selection matrix while regressing features onto the ground-truth labels. The consistent subspace representations together with the designed manifold regularization jointly guide the learning of feature selection matrix. An algorithm is established to implement the proposed model. Extensive experiments are conducted and experimental results demonstrate the superior performance of DSRSAG in comparison with representative PMFS methods.
Source: IEEE Transactions on Multimedia
Institution: Harbin Engineering University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,645
OpenAlex ID: W7168201671