A two-stage framework for discovering latent correlations in multi-label learning
Zijie Chen · 2016
It is a very important issue to discover data correlations in multi-label classification. A two-stage framework is presented to incorporate the supervised feature extraction and correlation exploration with the predictive modeling. Firstly, a low-dimensional feature mapping is obtained under the guidance of label information, and produces good feature extraction. Secondly, a predictive model is learnt on the extracted features. The proposed two-stage framework is efficient in low-dimensional problems. Furthermore, the dual form is presented to solve the high-dimensional problems more efficiently. Experiments show that the proposed framework achieves good performance on most datasets, especially when the correlations among data and labels are important. Besides, the framework is more efficient especially when the number of samples and the number of labels increase.