Large-Margin Label-Calibrated Support Vector Machines for Positive and Unlabeled Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Positive and unlabeled (PU) learning trains classifiers using only positive and unlabeled data.
• Existing PU learning methods often treat the problem as label noise or cost-sensitive learning.
• Current approaches often neglect data distribution, limiting performance.

Purpose of the Study:

• To propose a novel PU learning method that leverages data distribution information.
• To introduce a classifier that effectively utilizes clusters of positive and potential negative examples.
• To enhance the performance of PU learning models, especially when negative data is unavailable.

Main Methods:

• Developed a novel discriminative PU classifier named Large-margin Label-calibrated Support Vector Machines (LLSVM).
• Introduced a 'hat loss' to identify margins between data clusters.
• Incorporated a label calibration regularizer to correct biased decision boundaries.

Main Results:

• LLSVM effectively achieves a max-margin effect between positive and negative classes.
• The proposed method performs well even without explicit negative training examples.
• Theoretical analysis confirmed that PU data enhances algorithm performance.
• Empirical results demonstrated LLSVM's superiority over state-of-the-art PU methods on various datasets.

Conclusions:

• LLSVM offers a more effective approach to PU learning by considering data distribution.
• The method provides a robust solution for scenarios lacking negative training data.
• LLSVM advances the field of PU learning with its discriminative and label-calibrated approach.