The Forbidden Region Self-Organizing Map Neural Network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Self-Organizing Maps (SOMs) are unsupervised learning algorithms used for data visualization and clustering.
• Traditional SOMs do not account for 'forbidden regions' in the input space where samples cannot exist.
• Prototypes falling into these forbidden regions are meaningless and degrade map quality.

Purpose of the Study:

• To propose a novel SOM model that explicitly avoids placing prototypes in prespecified forbidden regions.
• To enhance the topological representation and vector quantization accuracy of SOMs when dealing with data containing such constraints.

Main Methods:

• A modified SOM algorithm was developed to incorporate constraints related to forbidden regions.
• The algorithm ensures that all generated prototypes remain outside these designated areas.
• The performance was evaluated against standard SOMs using experimental data sets.

Main Results:

• The proposed SOM model successfully kept all prototypes out of the specified forbidden regions.
• Experimental results demonstrated a lower vector quantization error compared to the standard SOM.
• The quality of the learned topological maps was significantly improved.

Conclusions:

• The new SOM model effectively handles data with forbidden regions, leading to more meaningful representations.
• This approach offers superior performance in both data representation accuracy and topological fidelity.
• The method is valuable for applications where prior knowledge about input space constraints is available.