Tri-Structured-Sparsity Induced Joint Feature Selection and Classification for Hybrid Noise Resilient Multilabel Learning

Abstract

Multilabel learning handles the problem that instances are associated with multiple labels. In practical applications, multilabel learning often suffers from imperfect training data. Typically, labels may be noisy or features may be corrupted, or both. Most existing multilabel learning models only consider either label noise or feature noise. Theoretically, ignoring any kind of noise in the learning process may lead to an unreasonable model, and thus affect the multilabel learning performance. In this paper, we propose a robust multilabel learning model, Tri-structured-Sparsity induced Joint Feature Selection and Classification (TriS-JFSC), to handle the data with hybrid noise. Specifically, the proposed TriS-JFSC model employs the tri-structured-sparsity regularization bridged with a label enhancement matrix to simultaneously smooth the feature and label noise, and embed a feature selection scheme that can simultaneously learn label-shared features and label-specific ones to boost the multilabel learning performance. Furthermore, by employing Alternating Direction Method of Multipliers (ADMM) method, a simple but efficient optimization algorithm is designed to solve the proposed TriS-JFSC model. Finally, the extensive experiments performed on several benchmark datasets demonstrate that our TriS-JFSC model outperforms other state-of-the-art learning methods.