Semi-supervised information-maximization clustering.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Semi-supervised clustering integrates prior knowledge into clustering algorithms.
• Existing unsupervised methods lack the ability to incorporate constraints like must-links and cannot-links.

Purpose of the Study:

• To propose a novel semi-supervised clustering algorithm based on the information-maximization principle.
• To extend a previous unsupervised algorithm to effectively incorporate must-links and cannot-links.

Main Methods:

• The proposed method is an extension of an unsupervised information-maximization clustering algorithm using squared-loss mutual information.
• It incorporates must-links and cannot-links into the clustering decision process.
• The clustering solution is obtained analytically via eigendecomposition, ensuring computational efficiency.

Main Results:

• The algorithm analytically incorporates must-links and cannot-links.
• It allows systematic optimization of tuning parameters, including kernel width.
• Experimental results demonstrate the usefulness and effectiveness of the proposed method.

Conclusions:

• The novel semi-supervised clustering algorithm provides an efficient and effective way to integrate prior knowledge.
• The analytical solution via eigendecomposition enhances computational performance.
• The method offers flexibility in parameter tuning based on the confidence in provided constraints.