Label distribution learning by utilizing common and label-specific feature fusion space

The article introduces a novel approach to Label Distribution Learning (LDL) called LDL-CLSFS, which leverages both common and label-specific features to enhance the performance of LDL models. Traditional LDL methods focus on shared features for all labels, but this approach often fails to capture the unique characteristics of each label. LDL-CLSFS addresses this limitation by constructing label-specific features that are most pertinent to each label. The method involves partitioning instances based on label-value rankings and employing clustering techniques to identify these features. These label-specific features are then combined with the common features to form a hybrid feature space, which is used to induce a label distribution model. The method is validated through extensive experiments on ten real-world datasets, demonstrating its superior performance compared to existing LDL methods. The article also discusses the importance of label-specific features and their potential to improve the performance of LDL models. Additionally, it highlights the flexibility and simplicity of the LDL-CLSFS method, which can be easily integrated with other LDL algorithms. The article concludes by noting the limitations of the current approach and suggesting future research directions, such as better integration of label-specific and common features, and the consideration of label correlations.