Multi-Label Classification by Label Clustering based on Covariance

Multi-label classification is a supervised learning problem that predicts multiple labels simultaneously. One of the key challenges in such tasks is modelling the correlations between multiple labels. LaCova is a decision tree multi-label classifier, that interpolates between two baseline methods: Binary Relevance (BR), which assumes all labels independent; and Label Powerset (LP), which learns the joint label distribution. In this paper we introduce LaCovaCLus that clusters labels into several dependent subsets as an additional splitting criterion. Clusters are obtained locally by identifying the connected components in the thresholded absolute covariance matrix. The proposed algorithm is evaluated and compared to baseline and state-of-the-art approaches. Experimental results show that our method can improve the label exact-match.