Multi-Label Active Learning Driven by Uncertainty and Inconsistency
Ran Wang, Suhe Ye · 2019
Multi-label active learning (MLAL) can help learn a highperformance multi-label classifier based on a smaller training set by selecting and labeling high-quality samples iteratively. This paper proposes a new MLAL algorithm by introducing a new concept called inconsistency cardinality. During each learning iteration, the current classifier can produce temporal decision labels for the candidate samples. Then, supporting decision regarding each label can be derived from the conditional features based on the lower approximations in fuzzy rough set theory. The inconsistency cardinality is defined as the disagreement degree between these two sets of decisions, which will be combined with uncertainty to give a more comprehensive evaluation on unlabeled samples. Correspondingly, a new MLAL strategy is proposed. Experimental comparisons show the feasibility and effectiveness of the proposed algorithm.