A Multi-label Learning Approach Based on Mapping from Instance to Label

Abstract

AbstractMulti-label classification approaches deal with ambiguous instances that may belong to several concepts simultaneously. In these learning frameworks, the inherent ambiguity of each instance is explicitly expressed in the output space where it is associated with multiple class labels. Recognizing the label sets for unseen instances becomes difficult because of the concept ambiguity. To handle with the multi-label learning problems, we propose a novel multi-label classification approach based on the assumption that, the relationship among instances in the feature space represents the relationship among their labels. We reconstruct a newly coming instance using the training data, and obtain a weight vector for it. This weight vector represents the relationship between the instance and the training instances, and its label vector can be obtained by the weighted sum of the label vectors of the training data. Experiments on real-world multi-label data sets show that, the approach achieves highly competitive performance compared with other well-established multi-label learning algorithms.KeywordsMulti-label learningClassificationSpace mapping