A unified multi-label classification framework with supervised low-dimensional embedding

作者：Chen, Zijie^*; Hao, Zhifeng

来源：Neurocomputing, 2016, 171: 1563-1575.

DOI：10.1016/j.neucom.2015.07.087

• Multi-label classification
• Data structure
• Label correlation
• Supervised low-dimensional embedding
• Dimensionality reduction
• Generalized eigenvalue problem

It is an important issue for multi-label classification to discover and utilize data structures or label correlations during the learning process, which could greatly improve the learning performance. In this paper, a unified framework is proposed for multi-label classification by incorporating the supervised low-dimensional embedding into the predictive model. The supervised embedding exploits latent structures and correlations from samples and labels, finds informative shared characteristics in a low-dimensional subspace and obtains a high quality dimensionality reduction. In the framework, a low-dimensional feature mapping is constructed through a linear transformation guided by the label information; meanwhile, the weights of the multi-label classifier have already been set up. The framework leads to a trace optimization problem and can be solved by a generalized eigenvalue problem. The dual form of the framework is also proposed to deal with high-dimensional cases. Experiments on ten datasets show that the proposed unified framework achieves better or comparable performance in terms of multi-label classification measures and ranking measures and needs much less training time in most cases. Furthermore, the framework is robust to the size of the low-dimensional subspace.

• 出版日期2016-1-1
• 单位华南理工大学; 广东药科大学; 广东工业大学