Dynamic Multi-label Learning with Multiple New Labels

Abstract

In a traditional multi-label learning task, an instance or object often has multiple labels. Previous works assume that the class labels are always fixed, i.e, the class labels in the test set are the same as that in the training set. Different from previous methods, we study a new problem setting where multiple new labels emerge in a dynamic environment. In this paper, we decompose the multiple labels pool to adjust the dynamic environment. The proposed method has several functions: classify instances on currently known labels, detect the emergence of several new labels then separate them using clustering, and construct a new classifier for each new label that works collaboratively with the classifier for known labels. Experimental results on publicly available data sets demonstrate that our method achieves superior performance, compared with the state-of-the-arts.