Multi-view Fusion and View-specific-labels Learning for Multi-label Classification

In multi-view multi-label learning, each sample is characterized by multiple views and simultaneously labeled with multiple semantic labels. Previous approaches typically combined information from various views to obtain a shared subspace, assuming each view has a shared set of labels. However, these methods ignore the diversity of multiple views, and each view can only capture a fraction of the complete label space. Moreover, the current approaches in this field often rely on either global label structures or local smoothness for learning label correlations, disregarding scenarios where both global and local correlations are crucial. To tackle these issues, we present a new method to improve multi-label classification through Multi-view Fusion and View-specific-labels Learning (MFVL). More concretely, we first get the shared subspace to capture shared information across the various views. Next, we combine it with the original multi-view dataset to construct a new multi-view dataset. To enhance multi-label classification performance, we learn view-specific labels and simultaneously consider global and local label correlations. Finally, our model is extended to nonlinear version to effectively handle linearly inseparable problem. In numerous experiments conducted on nine multi-view multi-label datasets, the proposed method achieves superior performance compared to state-of-the-art algorithms.