A Label Embedding Method for Multi-label Classification via Exploiting Local Label Correlations

Abstract

Multi-label learning has attracted more attention recently due to many real-world applications (e.g., text categorization and scene annotation). As the dimensionality of label space increases, it becomes more difficult to deal with this kind of applications. Therefore, dimensionality reduction techniques originally for feature space is also applied to label space, one of which is label embedding strategy which converts the high-dimensional label space into a low-dimensional reduced one. So far, existing label embedding methods mainly investigate the global recoverability between original labels and reduced labels, dependency between original features and reduced labels, or both. It is widely recognized that local label correlations could improve multi-label classification performance effectively. In this paper, we construct a trace ratio minimization problem as a novel label embedding criterion, which not only includes the global label recoverability and dependency, but also exploits the local label correlations as a local recoverability factor. Experiments on four benchmark data sets with more than 100 labels demonstrate that our proposed method is superior to four state-of-the-art techniques, according to two performance metrics for high-dimensional label space.