Statistical tools to assess the reliability of self-organizing maps

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Neural network learning, including self-organizing maps (SOM), exhibits inherent variability.
• This variability stems from factors like initial conditions, local minima convergence, and sampling differences.
• Assessing the reliability of SOM results is crucial for trustworthy data analysis.

Purpose of the Study:

• To develop and present a suite of tools for statistically evaluating the reliability of self-organizing map (SOM) outcomes.
• To provide a quantitative basis for assessing confidence in specific SOM results.
• To offer methods for evaluating the adequacy of the number of units in a SOM and comparing SOMs to other neural network models.

Main Methods:

• The proposed tools focus on analyzing the quantization error within a SOM.
• Neighborhood relations at both the individual observation pair level and the global map level are assessed.
• Statistical testing is employed to determine the confidence in SOM results.

Main Results:

• The developed tools enable a statistical assessment of the reliability of self-organizing map (SOM) results.
• Measures for quantization error and neighborhood relations provide insights into result stability.
• These methods also aid in determining the appropriate number of units for a SOM.

Conclusions:

• The presented tools offer a robust framework for validating self-organizing map (SOM) results.
• These reliability measures enhance the trustworthiness of SOM applications in data analysis.
• The tools can objectively demonstrate SOMs' reduced sensitivity to optimization issues compared to other neural networks.