Multi-Label Learning with Label-Specific Feature Selection via Local Positive and Negative Correlation

High dimension data increasingly emerged in many application areas. Feature selection methods are widely studied to solve the issues, playing a crucial role in eliminating redundant or irrelevant feature. Previous feature selection algorithms seldom consider the fact that each label might be determined by its specific features, which has its own characteristics. Meanwhile, they often neglect mutual exclusion among labels, named “negative label correlations” and assume labels share global correlation in exploring label correlation. To address the above problem, we propose a novel algorithm called “Multi-label learning with label-specific feature selection via local positive and negative correlation” (LLCPN-MIFS). Multiple datasets manifest that the effectiveness of the proposed method.