Multi-Label Classification With Label-Specific Feature Generation: A Wrapped Approach.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Multi-label classification involves assigning multiple labels to data instances.
• Existing methods often use a two-stage approach for label-specific feature generation and model induction, leading to suboptimal performance.
• Decoupling feature generation from model induction limits the exploitation of label correlations.

Purpose of the Study:

• To propose a novel wrapped learning approach for multi-label classification.
• To jointly optimize label-specific feature generation and classification model induction.
• To enhance the effectiveness of label-specific features in multi-label learning.

Main Methods:

• A wrapped learning strategy is proposed to integrate feature generation and model induction.
• Kernelized linear models are learned for each label.
• Label-specific features are simultaneously generated in an embedded feature space.
• Empirical loss minimization and pairwise label correlation regularization are employed.

Main Results:

• The proposed wrapped strategy effectively generates label-specific features.
• Joint optimization leads to improved multi-label classification performance.
• Comparative studies on sixteen benchmark datasets validate the approach's effectiveness.

Conclusions:

• The wrapped learning approach offers a more effective way to exploit label-specific features for multi-label classification.
• Jointly performing feature generation and model induction overcomes limitations of two-stage strategies.
• This integrated method demonstrates significant improvements across various datasets.