Label distribution learning based on ensemble neural networks

Abstract

Label distribution learning (LDL), as an extension of multi-label learning, is a new arising machine learning technique to deal with label ambiguity problems. The maximum entropy model is commonly used in label distribution learning. However, it does not consider the correlation between the labels and is not suitable for nonlinear relationships, and the prediction performance is also limited. In this paper, we propose a label distribution learning algorithm based on ensemble neural networks. The algorithm trains neural networks with preferences using training sets with different label sets to construct base learners, and combines the base learners with the weights, which is learned by the combined learner to obtain the final learning results. Experimental results show that the proposed algorithm is effective for label distribution data.