An Online Active Multi-label Classification Algorithm Based on a Hybrid Label Query Strategy
Kailun Gong, Tingting Zhai · 2021 3rd International Conference on Machine Learning, Big Data and Business Intelligence (MLBDBI) · 2021
Online active learning can both effectively reduce the labeling cost and process large-scale or streaming data, and thus it has become an important research area in machine learning. However, there are few online active learning studies regarding multi-label classification tasks. And the existing online active multi-label classification algorithms either ignore the label correlation and the label cardinality inconsistency, or can only be applied to a particular application field. In this paper, we propose a novel online active multi-label classification algorithm, termed MSGDA to overcome the drawbacks of the existing algorithms. Our hybrid label query strategy in MSGDA takes account of both the max-margin predictive uncertainty and the label cardinality inconsistency for measuring the importance of an unlabeled instance. Meanwhile, our online updating rule exploits the label correlation. The experimental results on six multi-label datasets demonstrate the superiority of our proposed algorithm.