Elastic K-means using posterior probability.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• K-means clustering is a prevalent hard clustering method.
• Existing methods often struggle with data exhibiting overlapping clusters or complex relationships.

Purpose of the Study:

• To propose Elastic K-means (EKM), a novel soft clustering algorithm.
• To integrate EKM with graph-based clustering for handling relational data.
• To establish theoretical underpinnings and demonstrate empirical effectiveness.

Main Methods:

• Developed Elastic K-means (EKM) utilizing posterior probabilities for fractional data point assignment.
• Integrated EKM with Normalized Cut graph clustering into a unified framework.
• Derived matrix inequalities to prove algorithm correctness and convergence.

Main Results:

• EKM enables data points to belong to multiple clusters fractionally.
• The integrated EKM-graph clustering model shows improved performance.
• Experimental validation on six benchmark datasets confirms the effectiveness of EKM and its integrated approach.

Conclusions:

• EKM offers a flexible and effective soft clustering alternative to hard clustering methods.
• The integration of EKM with graph clustering provides a powerful tool for complex data analysis.
• The theoretical framework supports the robustness and reliability of the proposed algorithms.