Robust low-rank representation with structured similarity learning for multi-label classification

Abstract

Handling high-dimensional, noisy data in multi-label classification is challenging, as feature abundance and noise obscure actual data-label relationships. Traditional approaches often model labels and features independently, limiting dependency modeling and noise reduction. To address this, we propose a unified framework combining low-rank representation using nuclear norm regularization with structured similarity learning. This simultaneously projects features and labels into low-rank spaces while preserving key inter-sample and inter-label relationships through structural constraints, further capturing fine-grained correlations via a learned similarity Matrix. Extensive experiments on five benchmark datasets show our model outperforms state-of-the-art methods, achieving a 16% reduction in Hamming Lossl and a 14% improvement in Micro-F1 on high-dimensional, noisy datasets like CAL500 and Corel16k7, with consistent gains in Macro-F1 and Example-F1. These results demonstrate the model’s strong capability for noisy, high-dimensional multi-label classification.