A Local and Global Discriminative Framework and Optimization for Balanced Clustering.

Area of Science:

• Data Mining
• Machine Learning
• Clustering Algorithms

Background:

• Many data mining and machine learning applications require balanced clusters, a challenge not met by existing algorithms.
• Current methods often fail to produce clusters of similar sizes or preserve natural data balance.

Purpose of the Study:

• To propose a novel balanced clustering framework utilizing both local and global data information.
• To develop methods that generate balanced clusters while preserving data structure and improving clustering quality.

Main Methods:

• Introduced Global Balanced Clustering (GBC) by combining a global partitioning model with distribution entropy minimization.
• Developed Global Balanced Regularization to adapt existing local clustering models for balanced outputs.
• Proposed Local and Global Balanced Spectral Clustering (LGB-SC) and Local and Global Balanced Local Learning (LGB-LL).
• Utilized augmented Lagrange multipliers for model optimization.

Main Results:

• The proposed framework successfully yields balanced clusters across various real-world benchmarks.
• The novel LGB-SC and LGB-LL models demonstrated superior performance compared to standard Spectral Clustering and Local Learning.
• The framework effectively balances cluster sizes while preserving high clustering quality.

Conclusions:

• The novel balanced clustering framework offers a significant advancement for applications requiring size-constrained partitions.
• The proposed LGB-SC and LGB-LL models provide effective solutions for balanced clustering, outperforming existing methods.