Batch and median neural gas.

Area of Science:

• Machine Learning
• Data Mining
• Computational Statistics

Background:

• Neural Gas (NG) is a robust clustering algorithm for Euclidean data, avoiding local minima and topological restrictions common in other methods like Self-Organizing Maps (SOM).
• Existing clustering algorithms often face limitations such as susceptibility to local minima or rigid topological constraints.

Purpose of the Study:

• To introduce a batch variant of the Neural Gas algorithm for enhanced performance.
• To develop a generalized median-based variant of Neural Gas for non-vectorial proximity data.
• To provide a unified framework for proving the convergence of batch and median versions of NG, SOM, and k-means.

Main Methods:

• Development of a batch variant of Neural Gas based on its cost function, interpreted as Newton method optimization.
• Introduction of a non-vectorial data variant using the generalized median concept, analogous to Median SOM.
• Unified theoretical formulation to prove convergence for batch and median NG, SOM, and k-means algorithms.

Main Results:

• The batch variant of Neural Gas demonstrates significantly faster convergence compared to the standard version.
• The generalized median variant extends Neural Gas capabilities to handle non-vectorial proximity data effectively.
• Experimental investigations validate the convergence and behavior of the proposed algorithms.

Conclusions:

• The batch and median variants of Neural Gas offer substantial improvements in convergence speed and data applicability.
• The unified convergence proof provides a strong theoretical foundation for these clustering algorithms.
• These advancements contribute to more robust and versatile clustering solutions in machine learning.