Online Multi-Label Classification with Scalable Margin Robust to Missing Labels

Abstract

Multi-label classification with missing labels handles the problem that the label set contains unobserved missing labels due to the expensive human annotations. However, these works mainly concentrate on offline settings, and thus fail to deal with online streaming data in real time. In this paper, we focus on online multi-label learning with missing labels, and develop a novel maximum margin-based online algorithm with scalable margin (OML-SM). Specifically, we update the margin-based model incrementally based on the past model and the current incoming instance with its partially observed labels. To build robustness to missing labels, we first tune the classification margin of the negative labels via the label causality matrix, which is formed by the conditional probability of label pairs. Secondly, the label prototype matrix is proposed to adjust the margin by controlling the balancing parameter of the slack term. To adapt to the possible evolution of label distribution, the aforementioned two matrices are also incrementally updated based on the partially observed label set in an efficient way. Finally, the validity of the proposed method on multiple evaluation metrics is further supported by experiments conducted on a number of benchmark data sets.