Improving the performance of self-organizing maps via growing representations

Area of Science:

• Artificial Intelligence
• Computational Chemistry
• Bioinformatics

Background:

• Self-organizing maps (SOMs) are artificial neural networks for simplifying high-dimensional data.
• Traditional SOMs use static node configurations and are applied in pharma and bioinformatics.
• Existing SOMs are limited by fixed connectivity and topology.

Purpose of the Study:

• Investigate the impact of connectivity and topology on SOM performance.
• Compare static and growing SOMs using synthetic and chemistry domain datasets.
• Assess performance based on topological and quantization errors.

Main Methods:

• Implemented and evaluated fixed and growing SOM architectures.
• Utilized three synthetic and two chemistry datasets for experiments.
• Assessed SOM performance using topological and quantization error metrics.

Main Results:

• All SOMs showed comparable data space quantization.
• Growing SOMs significantly outperformed static SOMs in capturing data structure (topological errors).
• The Neural Gas model demonstrated superior ability in identifying hidden low-dimensional relationships.

Conclusions:

• Growing SOMs offer advantages over static configurations for data structure representation.
• The Neural Gas model is particularly effective for uncovering complex relationships in high-dimensional data.
• Topology and connectivity significantly influence SOM performance in capturing underlying data patterns.