A unified view of density-based methods for semi-supervised clustering and classification.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• The increasing volume of big data exacerbates the challenge of limited labeled datasets.
• Semi-supervised learning addresses this by utilizing both labeled and unlabeled data.
• Density-based clustering and graph-based transductive classification are prominent techniques.

Purpose of the Study:

• To introduce a unified perspective on density-based clustering algorithms.
• To bridge semi-supervised clustering and classification using density-based methods.
• To develop a novel, statistically sound framework for semi-supervised classification and clustering.

Main Methods:

• A unified view of density-based clustering algorithms is presented.
• Connections between density-based clustering and graph-based transductive classification are established.
• A new framework for semi-supervised classification is developed using density-based clustering components.
• The HDBSCAN* algorithm is generalized for semi-supervised clustering.

Main Results:

• The proposed framework demonstrates efficiency and effectiveness in semi-supervised classification.
• The generalized HDBSCAN* algorithm successfully performs semi-supervised clustering.
• Experimental results confirm the advantages of the approach across various datasets.

Conclusions:

• Density-based techniques provide a robust foundation for unifying semi-supervised clustering and classification.
• The developed framework offers a statistically sound and efficient solution for semi-supervised learning tasks.
• The generalized HDBSCAN* algorithm enhances the utility of density-based methods for semi-supervised clustering.