Semantic-Gap-Oriented Feature Selection and Classifier Construction in Multilabel Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Multilabel learning assigns multiple labels to instances, aiming to map feature space to label space.
• This process involves both feature selection and classifier construction.
• A key challenge is the semantic gap between input features and output labels.

Purpose of the Study:

• To develop a robust multilabel learning framework that addresses the semantic gap.
• To enhance feature selection and classifier construction by leveraging feature-label correlations.
• To improve the accuracy of predictive functions in multilabel classification.

Main Methods:

• Feature selection based on learned positive and negative feature-label correlations.
• Sample-specific and label-specific classifications incorporating interlabel and interinstance correlations.
• Bridging the semantic gap using learned feature-label correlations.

Main Results:

• The proposed framework effectively extracts label-specific features.
• It successfully performs sample- and label-specific classifications.
• Experimental results on multiple benchmarks across four domains validate the framework's effectiveness.

Conclusions:

• The developed framework offers an effective solution for multilabel learning.
• It successfully bridges the semantic gap, improving predictive function learning.
• The approach demonstrates strong performance on diverse datasets.