Dual Space Latent Representation Learning for Image Representation

Semi-supervised non-negative matrix factorization (NMF) has achieved successful results due to the significant ability of image recognition by a small quantity labeled information. However, there still exist problems be solved such as interconnection information not being fully explored and inevitable mixed noise in data, which deteriorates performance these methods. To circumvent this problem, we propose novel semi-supervised method named DLRGNMF. Firstly, dual latent space is characterized affinity explicitly reflect interrelationship between data instances feature variables, can exploit global reduce adverse impacts caused redundant Secondly, embed manifold regularization mechanism graph steadily retain local structure space. Moreover, sparsity biorthogonal condition are integrated constrain factorization, greatly improve algorithm’s accuracy robustness. Lastly, an effective alternating iterative updating proposed, model optimized. Empirical evaluation on nine benchmark datasets demonstrates that DLRGNMF more than competitive