Multi-Label Learning Based on Label Entropy Guided Clustering

Abstract

Recently multi-label learning has attracted the attention of a lot of researchers in machine learning field. Many algorithms have been proposed. The main stream of multi-label learning is the research on how to boost predicting performance using label correlations. However, these methods ignore the importance of feature vectors. Recent study explores to use feature vectors and label vectors collaboratively. This paper proposes a simple but effective algorithm ML-LEC (Multi-label Learning based on Label Entropy guided Clustering). It first performs clustering with the number of clusters set by label entropy adaptively for each label. New features are constructed from the original feature vectors by querying the clustering result. Then, models are obtained by using ordinary classification algorithm. Experiments on several data sets from different application domains verify the superiority of the proposed algorithm to some baseline and the state-of-art ones.