Multi-label Classification: A Comparative Study on Threshold Selection Methods

Dealing with multiple labels is a supervised learning problem of increasing importance. However, in some tasks, certain learning algorithms produce a confidence score vector for each label that needs to be classified as relevant or irrelevant. More importantly, multi-label models are learnt in training conditions called operating conditions, which most likely change in other contexts. In this work, we explore the existing thresholding methods of multi-label classification by considering that label costs are operating conditions. This paper provides an empirical comparative study of these approaches by calculating the empirical loss over range of operating conditions. It also contributes two new methods in multilabel classification that have been used in binary classification: score-driven and one optimal.