Learning to reweight examples in multi-label classification.

Abstract

This paper presents a new method for reweighting examples in the multi-label classification problem. Existing weighting functions in self-paced learning simply determine the weights of examples according to their loss values given the current multi-label model, but neglect the unique properties of multi-label examples. It is inaccurate to treat two distinct examples as equal even if their loss values are the same. Therefore, we upgrade the classical weight functions by considering instance complexities, which are described by the distances between instance features and their corresponding labels. The distance metric can be easily optimized during training. Experimental results on real-world datasets demonstrate the significance of investigating both the dynamic and static complexities of multi-label examples, as well as the advantages of the proposed example reweighting algorithm in multi-label classification problems.