Label-specific feature selection and two-level label recovery for multi-label classification with missing labels

Abstract In multi-label learning, each instance is assigned by several nonexclusive labels. However, these labels are often incomplete, resulting in unsatisfactory performance in label related applications. We design a two-level label recovery mechanism to perform label imputation in training sets. An instance-wise semantic relational graph and a label-wise semantic relational graph are used in this mechanism to recover the label matrix. These two graphs exhibit a capability of capturing reliable two-level semantic correlations. We also design a label-specific feature selection mechanism to perform label prediction in testing sets. The local and global feature-label connection are both exploited in this mechanism to learn an inductive classifier. By updating the matrix that represents the relevance between features and the predicted labels, the label-specific feature selection mechanism is robust to missing labels. At last, intensive experimental results on nine datasets under different domains are presented to demonstrate the effectiveness of the proposed approach.