Frequency-sensitive competitive learning for scalable balanced clustering on high-dimensional hyperspheres.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• High-dimensional data clustering, especially with competitive learning, faces challenges due to the curse of dimensionality.
• Spherical k-means (spkmeans) is effective for normalized high-dimensional data but often yields imbalanced clusters.
• Existing methods struggle with generating balanced clusters for large numbers of clusters in high-dimensional spaces.

Purpose of the Study:

• To develop novel frequency-sensitive competitive learning algorithms for balanced high-dimensional clustering.
• To adapt a maximum likelihood formulation using von Mises-Fisher distributions for improved clustering.
• To create scalable algorithms for both static and streaming high-dimensional data.

Main Methods:

• Derived spkmeans from a maximum likelihood formulation with a von Mises-Fisher mixture model.
• Adapted the generative model to create three frequency-sensitive competitive learning variants for static data.
• Developed a frequency-sensitive algorithm for clustering streaming data.

Main Results:

• The proposed frequency-sensitive competitive learning variants produce high-quality and well-balanced clusters for high-dimensional data.
• All proposed algorithms exhibit linear time complexity per iteration concerning data points and clusters.
• Experimental results demonstrate the effectiveness on high-dimensional text datasets.

Conclusions:

• Frequency-sensitive competitive learning offers a principled approach to achieve balanced clustering in high-dimensional spaces.
• The developed methods effectively address the limitations of spkmeans regarding cluster balance.
• The techniques are scalable and applicable to both static and streaming high-dimensional data clustering tasks.