Auto-SOM: recursive parameter estimation for guidance of self-organizing feature maps.

Area of Science:

• Computational intelligence
• Machine learning
• Data visualization

Background:

• Kohonen's self-organizing map (SOM) is vital for dimensionality reduction and exploratory data analysis.
• Traditional SOMs require manual, heuristic parameter tuning for optimal performance.
• Neighborhood preservation is crucial for accurate low-dimensional representations.

Purpose of the Study:

• To introduce Auto-SOM, an algorithm for automatic learning parameter estimation in SOMs.
• To enable user-defined control over neighborhood violation degrees in mappings.
• To enhance the neighborhood preservation capabilities of self-organizing maps.

Main Methods:

• Auto-SOM integrates a Kalman filter-based SOM with recursive parameter estimation.
• Kalman filter optimizes neuron weights using estimated learning coefficients to minimize estimation error variance.
• Neighborhood function width is estimated by minimizing Kalman filter prediction error variance, incorporating a topographic function to prevent violations.

Main Results:

• Auto-SOM successfully automates the estimation of critical learning parameters for SOMs.
• The algorithm effectively prevents neighborhood violations to a user-specified degree in both mapping directions.
• Demonstrated preservation of neighborhood structures in low-dimensional mappings through three application examples.

Conclusions:

• Auto-SOM significantly improves the usability and performance of self-organizing maps.
• This automated approach facilitates the creation of reliable neighborhood-preserving maps for dimension reduction.
• The method offers a robust solution for exploratory data analysis requiring accurate neighborhood preservation.