Learning tree structure of label dependency for multi-label learning

There always exists some kind of label dependency in multi-label data. Learning and utilizing those dependencies could improve the learning performance further. Therefore, an approach for multi-label learning is proposed in this paper, which quantifies the dependencies of pairwise labels firstly, and then builds a tree structure of the labels to describe them. Thus the approach could find out potential strong label dependencies and produce more generalized dependent relationships. The experimental results have validated that compared with other state-of-the-art algorithms, the method is not only a competitive alternative, but also has shown better performance after ensemble learning especially. Copyright © 2012 Springer-Verlag Berlin Heidelberg.