Instance-Dependent Positive and Unlabeled Learning with Labeling Bias Estimation.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Positive and Unlabeled (PU) classification presents challenges due to non-uniform labeling probabilities.
• Existing methods often assume instance-independent labeling, which may not reflect real-world data complexities.
• Instance-dependent labeling, where labeling depends on both class and observation, requires novel approaches.

Purpose of the Study:

• To develop a novel framework for instance-dependent PU classification.
• To accurately model the relationship between instance features, class labels, and labeling probability.
• To improve the performance of PU classification algorithms by accounting for instance-specific labeling biases.

Main Methods:

• A graphical model is constructed to represent the instance-dependent labeling probability P(s,y|x).
• A maximization problem is formulated based on the induced likelihood function.
• Expectation-Maximization (EM) and Adam optimization techniques are employed to estimate labeling probabilities and classifiers.

Main Results:

• The proposed method successfully estimates instance-specific labeling probabilities P(s=1|y=1,x) and classifiers P(y|x).
• Theoretical analysis confirms the existence and local uniqueness of critical solutions for linear models under sufficient conditions.
• Generalization error bounds are established for both linear logistic and non-linear network implementations.
• Empirical evaluations on diverse datasets demonstrate superior performance compared to state-of-the-art PU algorithms.

Conclusions:

• The proposed instance-dependent PU classification method effectively addresses labeling biases.
• The framework provides a robust theoretical foundation and practical advantages over existing approaches.
• This work advances the field of PU learning with significant improvements in accuracy and applicability.