Robust Semi-supervised Multi-label Learning by Triple Low-Rank Regularization

Abstract

Multi-Label Learning (MLL) deals with the problem when one instance is associated with multiple labels simultaneously. Previous methods have shown promising performance by effectively exploiting the semantic correlations among different labels. However, most of the existing methods may not be robust to the situation when the training instances are labeled with noisy or incomplete labels, which are common in reality. In this paper, we propose Robust Semi-Supervised Multi-Label Learning by Triple Low-Rank Regularization approach to address this problem. Specifically, a linear self-representative model is firstly introduced to recover the possibly noisy label matrix by exploiting the label correlations. Then, our method develops a low-rank pairwise similarity matrix to capture the global relationships among labeled and unlabeled samples by taking advantage of Low-Rank Representation (LRR). In addition, by utilizing the pairwise similarity matrix defined above, we construct the graph Laplacian regularization to acquire geometric structural information from both labeled and unlabeled samples. Moreover, the proposed method concatenate the prediction models for different labels into a matrix, and introduces the matrix trace norm to capture the correlations and control the model complexity. Experimental studies across a wide range of benchmark datasets show that our method achieves highly competitive performance against other state-of-the-art approaches.