Cluster-infused low-rank subspace learning for robust multi-label classification

Abstract

Multi-label learning in high-dimensional spaces Suffers from the curse of dimensionality, noisy labels, and complex feature-label dependencies. Traditional deep learning solutions for multi-label classification employ multi-layer networks but overfit and generalize poorly owing to ineffective high-order data dependencies. In this paper, we introduce a cluster-infused low-rank subspace learning framework that integrates low-rank subspace learning with cluster infusion to solve these issues. Our model resolves sensitivity to noise, overfitting and poor generalization in high-dimensional data by using low-rank subspace representation decomposition of the classifier for dimension reduction and low-rank classifier for discriminative classification. To enhance robustness, we reconstruct each data sample as a Linear combination of its neighbours, infusing clustering-derived features into the model. These facilitate feature robustness via local correlations, thereby improving noise resilience and discriminative power. Extensive experiments on benchmark high-dimensional datasets, compared against state-of-the-art approaches, indicate that our approach significantly improves classification accuracy and robustness, making it a good solution for noisy, high-dimensional multi-label classification tasks. This effectiveness is evidenced across datasets of various scales, including a 3.04% improvement in Example-F1 over CNN-RNN on the smaller 20NG dataset and a significant 9.9% gain in Micro-F1 against RethinkNet on the large-scale NUS-WIDE dataset, highlighting DL-CS’s superiority for diverse multi-label classification tasks.