Positive-Unlabeled Learning With Label Distribution Alignment.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Positive-Unlabeled (PU) data presents challenges for binary classification due to the lack of negative labels.
• Existing PU learning methods often exhibit a bias towards negative predictions, limiting their effectiveness.

Purpose of the Study:

• To propose a novel formulation for PU learning that alleviates the bias towards negative predictions.
• To introduce a margin-based learning framework, PULDA, enhancing model discriminability and controlling negative prediction proportions.

Main Methods:

• Developed a label distribution alignment formulation to control the proportion of negative predictions globally.
• Introduced functional margins to improve the discriminability of the classifier.
• Integrated class prior estimation and a stochastic mini-batch optimization algorithm with convergence guarantees.

Main Results:

• The proposed label distribution alignment intrinsically eliminates bias towards negative predictions.
• The PULDA framework demonstrates enhanced model discriminability.
• Empirical results confirm the effectiveness of the proposed method in PU learning scenarios.

Conclusions:

• PULDA offers an effective solution for PU learning by addressing inherent biases and improving classification performance.
• The method is theoretically supported by generalization analysis and practically validated through comprehensive experiments.