Confidence-Based PU Learning With Instance-Dependent Label Noise.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Positive and Unlabeled (PU) learning trains classifiers using only PU data.
• Traditional PU learning assumes accurate positive labels, which is often not true in practice.
• Label noise in the positive set is common and can be instance-dependent.

Purpose of the Study:

• To address the understudied problem of PU learning with instance-dependent label noise (PUIDN).
• To develop a method that mitigates the adverse impact of noisy positive labels without assuming noise distribution.
• To improve the robustness and accuracy of PU learning algorithms.

Main Methods:

• Leveraging confidence scores for each instance in the positive set.
• Proposing an unbiased estimator for classification risk using label and confidence information.
• Integrating an alternating iteration optimization strategy based on confidence correlations.

Main Results:

• The proposed method effectively handles instance-dependent label noise in PU learning.
• An unbiased risk estimator is developed and computed from PUIDN data.
• The framework demonstrates improved performance through experimental validation.

Conclusions:

• The developed approach offers a robust solution for PU learning with instance-dependent label noise.
• Confidence scores are crucial for connecting samples and labels in noisy scenarios.
• The method provides a theoretical generalization error bound and practical effectiveness.