Label Enhancement with Sample Correlations via Low-dimensional Feature Representation

In this paper we propose a novel label enhancement (LE) algorithm called Label Enhancement with sample correlations via Low-dimensional Feature Representation (LE-LFR) in order to solve the problem that many training sets cannot use label distribution learning (LDL) algorithms because they only contain logical labels rather than label distributions. Unlike most existing two-stage methods, LE-LFR method is a one-stage method and the sample correlations are mined comprehensively. In detail, we obtain the low-dimensional feature representation via a manifold learning process firstly and then we get the label distribution during the intermediate procedure via the label propagation process. The sample correlation both in the label space and low-dimensional feature space is considered during the process. Finally, the enhanced maximum entropy model is established as objective function, and the predicted label distribution is obtained. By using the rich label information in low-dimensional feature space obtained in the first step and the label distribution information estimated in the second step, the enhanced maximum entropy prediction model is trained through gradient descent iterative optimization, and the label distribution prediction parameters with higher accuracy are obtained. Experimental results on fourteen real-world datasets show superior advantages of LE-LFR against several existing LE algorithms.