Learning From PU Data Using Disentangled Representations.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Classical Positive Unlabeled (PU) learning methods struggle with high-dimensional data complexity.
• Existing high-dimensional PU learning techniques are also affected by data complexity.

Purpose of the Study:

• To develop a robust PU learning method for high-dimensional data.
• To enhance the efficacy of clustering techniques in complex datasets.
• To improve the identification of positive and negative classes in partially labeled data.

Main Methods:

• Utilized a neural network to learn a data representation.
• Employed a novel loss function to project unlabeled data into distinct positive and negative clusters.
• Implemented a vector quantization strategy to refine cluster separation.

Main Results:

• Demonstrated superior performance over state-of-the-art methods on benchmark PU datasets.
• Successfully projected unlabeled data into well-separated positive and negative clusters.
• Validated the effectiveness of the neural network-based representation learning.

Conclusions:

• The proposed cluster-based neural network approach effectively addresses high-dimensional PU learning challenges.
• The method offers a simplified approach to PU learning, akin to low-dimensional settings.
• Theoretical justification supports the cluster-based strategy and algorithmic choices for enhanced classification.