UC San Diego

Multi-label classification is widely used for various applications such as automatic music tagging. Often, multi -label learning is done by transforming into multiple independent binary classification problems. In order to produce better classification result, label correlations should be taken into account. This thesis first discusses how to model label correlations in a quantitative way and categorizes the concept into unconditional and conditional correlations. After that, this thesis shows how to exploit both kinds of label correlations for multi-label learning algorithms. The main model this thesis addresses is conditional random fields (CRFs). This thesis shows how to apply CRFs for multi-label classification. Because of the intractable nature of CRF inference, several approximation algorithms to make it applicable for larger label sets are described. Various other learning algorithms that exploit label correlations are also discussed in this thesis. In the end, all the mentioned multi-label learning algorithms are evaluated with a music data set, CAL500, composed of 502 songs categorized into 174 labels