Multi-Label Learning With Label Correlation Based On Rbf Network

In the real world, things always have a variety of characteristics, so in machine learning, we need multi-label to describe this situation. Nowadays, multi-label problem is the widely used in modern applications. While different from traditional classification, multi-label problem is more difficult to accomplish, more and more efforts have been put into the research of this problem.During the past decade, a large amount of algorithms which learns from multi-label data have been proposed, most of them focus on manipulating with identical feature set However, this popular strategy might be suboptimal as each label is supposed to possess specific characteristics of its own. What's more, in multi-label problems, each label is not independent to others and label correlation is very important to improve the accuracy of classification, so our target is trying to build specific features of each label and to exploit label correlation for improving the accuracy performance of multi-label learning.In order to build specific features for learning, a RBF network is introduced to multi-label learning problem for mapping original identical features to specific features for each label, this operation makes model of each label learns from specific characteristics of its own. Moreover, a novel approach using predicted labels which has prunes error-prone ones as additional features for learning label correlations automatically. According to the experiment on the several benchmark data sets of multi-label problems, we have verified that our proposed method is effective in improving the accuracy of multi-label problem classification, meanwhile the running time of proposed method is acceptable.