Multi-Label Low-dimensional Embedding with Missing Labels

Multi-label learning is to predict proper label sets for each training sample. Usually, the label sets of instances within the same group exhibit certain similarities. It means label sets sharing the same cluster are strongly correlated with each other while label sets of other clusters are loosely correlated. In this study, we calculate the instance-wise cosine similarity on label sets of three multi-label benchmarks in different applications to validate our hypothesis. To facilitate label imputation procedure, we exploit the low rank and sparse properties to capture the global structure of label sets in instance level. Besides, some datasets may not show clear separation of label sets by topics. The proposed label recovery method can also handle this kind of datasets. In addition to the instance-wise label correlation used in the output space to handle missing labels, the feature and label connection is also mined in the input space to learn an inductive classifier for out-of-sample extrapolation. Experimental results on three benchmark datasets in image annotation, action units detection and text categorization demonstrate the effectiveness of the proposed method on datasets with distinct category; experimental results on nine benchmark datasets in music, biology, video, audio, image and text domains demonstrate the effectiveness of the proposed method on datasets with indistinct category.