Multilabel Classification With Label-Specific Features and Classifiers: A Coarse- and Fine-Tuned Framework.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Multilabel classification involves assigning multiple labels to instances simultaneously.
• Existing methods struggle to effectively integrate feature selection and classifier training.
• Learning a robust mapping from feature to label space is crucial for out-of-sample extrapolation.

Purpose of the Study:

• To develop a unified mechanism for extracting label-specific features and training label-specific classifiers.
• To enhance the robustness of mapping learning in multilabel classification by combining local and global feature extraction with individual and joint classifier training.
• To address the limitations of current approaches in effectively learning the mapping function across feature and label domains.

Main Methods:

• Derivation of mechanisms for label-specific feature extraction at local and global levels.
• Derivation of mechanisms for label-specific classifier training at individual and joint levels.
• Implementation of a two-level feature selection and two-level classifier training strategy for a robust mapping learning process.

Main Results:

• Extensive experimental results demonstrate the effectiveness of the proposed approach.
• The method shows significant improvements on several benchmark datasets across four domains.
• The dual-level approach (coarse-tuned global feature extraction/joint classifier training and fine-tuned local feature extraction/individual classifier training) proved robust.

Conclusions:

• The proposed method effectively learns the mapping function in multilabel classification by integrating feature extraction and classifier training.
• The two-level feature selection and classifier training enhance the robustness and performance of multilabel classification models.
• The approach offers a significant advancement in handling complex data with multiple labels.