Positive Unlabeled Learning Selected Not At Random (PULSNAR): class proportion estimation without the selected...

Area of Science:

• Machine Learning
• Data Science
• Computer Science

Background:

• Positive and unlabeled (PU) learning is a semi-supervised classification method crucial when obtaining negative instances is difficult.
• Existing PU algorithms often assume positives are selected completely at random (SCAR), which fails in many real-world applications like healthcare.
• The SCAR assumption violation leads to inaccurate estimation of the positive proportion (α) and poor model calibration.

Purpose of the Study:

• To develop novel PU learning algorithms that can estimate α and provide calibrated probabilities for unlabeled instances.
• To improve classification metrics in scenarios where positives are selected not at random (SNAR).

Main Methods:

• Proposed two algorithms: PULSCAR (Positive Unlabeled Learning Selected Completely At Random) and PULSNAR (Positive Unlabeled Learning Selected Not At Random).
• PULSNAR utilizes a divide-and-conquer strategy to cluster SNAR positives into subtypes.
• α is estimated for each subtype by applying PULSCAR to the clustered positives and all unlabeled data.

Main Results:

• PULSNAR effectively handles SNAR data, improving the estimation of α and model calibration.
• The proposed algorithms demonstrated superior performance compared to state-of-the-art methods on synthetic and real-world datasets.
• PULSNAR achieved better classification metrics, indicating more reliable identification of positive instances.

Conclusions:

• The PULSNAR algorithm offers a robust solution for PU learning problems with SNAR characteristics.
• This approach enhances the practical applicability of PU learning in fields like healthcare and drug discovery.
• The developed methods provide more accurate and reliable classification in challenging semi-supervised settings.