Extracting Privileged Information for Enhancing Classifier Learning.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Data-driven learning accuracy is limited by insufficient or poor-quality training data.
• Manually labeled privileged information (PI) improves classifiers but is time-consuming and may lack richness.

Purpose of the Study:

• To enhance classifier learning by automatically exploring rich PI from untagged corpora.
• To eliminate the dependency on manual labeling for privileged information.
• To develop a more powerful category classifier by integrating subcategory classifiers.

Main Methods:

• Treating each selected PI as a subcategory and learning independent classifiers.
• Integrating subcategory classifiers to form a powerful category classifier.
• Proposing a novel instance-level multi-instance learning (MIL) model for simultaneous image selection and SVM classifier optimization.

Main Results:

• The proposed approach effectively enhances classifier learning by leveraging automatically explored PI.
• Experiments on four benchmark datasets demonstrate the superiority of the proposed method.
• The MIL model successfully selects training images and optimizes SVM classifiers.

Conclusions:

• Exploring PI from untagged corpora is a viable strategy to improve classifier accuracy.
• The novel MIL model offers an effective solution for integrating automatically derived PI.
• The approach significantly outperforms existing methods in improving data-driven learning accuracy.