Weighted Contrastive Learning With Hard Negative Mining 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.
• Current PU learning methods often treat unlabeled data as negative, leading to poor data representations and performance.
• Existing approaches struggle with generating high-quality data representations, causing negative-prediction bias and reduced accuracy.

Purpose of the Study:

• To propose a novel algorithm, weighted contrastive learning with hard negative mining for positive and unlabeled learning (WConPU), to address limitations in PU learning.
• To enhance the quality of data representations for improved PU learning performance.
• To provide theoretical justification for the proposed method using the expectation-maximization (EM) algorithm framework.

Main Methods:

• Developed a novel algorithm, WConPU, integrating a contrastive learning (CL) module and a classifier training module.
• Designed a new prototypical contrastive strategy for discriminative representation learning in PU learning.
• Introduced a weighted contrastive objective function with a prototype-based hard negative mining module to boost representation quality.

Main Results:

• The proposed WConPU algorithm iteratively enhances representation quality and classifier performance.
• Theoretical analysis demonstrates the method's alignment with the expectation-maximization (EM) algorithm.
• Empirical evaluations on diverse benchmark datasets show WConPU outperforms state-of-the-art PU learning methods.

Conclusions:

• WConPU effectively generates discriminative representations for PU learning tasks.
• The algorithm overcomes the negative-prediction preference and performance decline seen in existing methods.
• WConPU offers a significant advancement in positive and unlabeled learning methodologies.