Multi-label Active Learning with Error Correcting Output Codes

Due to the demand of practical problems, multi-label learning has become an important research where each instance belongs to multiple classes. Compared with single-label problem, the labeling cost for multi-label one is rather expensive because of the diversity and non-uniqueness of the labels. Therefore, the active learning which reduces the cost by selecting the most valuable data to query the labels attracts a lot of interests. Although several multi-label active learning (MLAL) methods were proposed, they often identify the label merely through a classifier via one-versus-all (OVA) strategy for each class, which makes the classification model very fragile, thus having a serious impact on the later selection criteria. In this paper, we utilize a new multi-label Error Correcting Output Codes (ECOC) method which determines the label of an instance on each class by combining multiple classifiers. This makes our classification model has a good ability of error-correcting and thus ensures the effectiveness of evaluation information in the selection process. Then we combine two effective selection strategies, the margin prediction uncertainty and label cardinality inconsistency, to complement each other and select the most informative instance. Based on this combination, we propose a novel MLAL framework, termed Multi-label Active Learning with Error Correcting Output Codes (MAOC). Experiments on multiple benchmark multi-label datasets demonstrate the efficacy of the combination in proposed approach.