Auxiliary Label Embedding for Multi-label Learning with Missing Labels

Label correlation has been exploited for multi-label learning in different ways. Existing approaches presume that label correlation information is available as a prior, but for multi-label datasets having incomplete labels, the assumption is violated. In this paper, we propose an approach for multi-label classification when label details are incomplete by learning auxiliary label matrix from the observed labels, and generating an embedding from learnt label correlations preserving the correlation structure in model coefficients. The approach recovers missing labels and simultaneously guides the construction of model coefficients from the learnt label correlations. Empirical results on multi-label datasets from diverse domains such as image & music substantiate the correlation embedding approach for missing label scenario. The proposed approach performs favorably over four popular multi-label learning techniques using five multi-label evaluation metrics.