# self-organizing map

> machine learning technique useful for dimensionality reduction

**Wikidata**: [Q1136838](https://www.wikidata.org/wiki/Q1136838)  
**Wikipedia**: [English](https://en.wikipedia.org/wiki/Self-organizing_map)  
**Source**: https://4ort.xyz/entity/self-organizing-map

## Summary
A self-organizing map (SOM) is a machine learning technique used for dimensionality reduction that maps high-dimensional data onto a lower-dimensional grid while preserving topological relationships. It is a type of artificial neural network that organizes data into clusters based on similarity.

## Key Facts
- SOMs are a type of artificial neural network used for unsupervised learning
- They are also known as Kohonen networks, named after their inventor Teuvo Kohonen
- SOMs reduce high-dimensional data to typically 2D or 3D representations
- The algorithm preserves the topological properties of the input space
- SOMs are commonly used for visualization, clustering, and data mining applications
- The network consists of neurons arranged in a grid (usually 2D) that compete to represent input data
- Training involves presenting input vectors to the network and adjusting weights to minimize distance
- SOMs are particularly effective for pattern recognition and exploratory data analysis

## FAQs
### Q: What is a self-organizing map used for?
A: Self-organizing maps are used for dimensionality reduction, data visualization, clustering, and pattern recognition. They help transform complex high-dimensional data into simpler 2D or 3D representations while preserving the relationships between data points.

### Q: How does a self-organizing map work?
A: A SOM works by creating a grid of neurons that compete to represent input data. When an input vector is presented, the neuron closest to it (best matching unit) is activated, and its weights are adjusted along with those of neighboring neurons. This process repeats iteratively, organizing the network to reflect the structure of the input data.

### Q: What makes SOMs different from other neural networks?
A: Unlike supervised neural networks, SOMs are unsupervised and focus on organizing data rather than learning specific input-output mappings. They preserve topological relationships and create a spatial organization of data where similar items are mapped to nearby locations on the grid.

## Why It Matters
Self-organizing maps are significant because they provide an effective way to visualize and understand complex high-dimensional data. In an era where datasets often contain hundreds or thousands of variables, SOMs offer a powerful tool for reducing this complexity to human-interpretable 2D or 3D representations. This capability makes them invaluable for exploratory data analysis, helping researchers and analysts identify patterns, clusters, and relationships that would be difficult to discern in the original high-dimensional space. SOMs have found applications across diverse fields including bioinformatics, text mining, image analysis, and market segmentation, where understanding the structure of complex data is crucial for decision-making and discovery.

## Notable For
- Being one of the first neural network architectures specifically designed for unsupervised learning
- Successfully preserving topological relationships during dimensionality reduction
- Providing intuitive visual representations of complex data structures
- Being computationally efficient compared to other dimensionality reduction techniques
- Having a simple yet powerful architecture that requires minimal parameter tuning

## Body
Self-organizing maps were developed by Finnish professor Teuvo Kohonen in the 1980s as a neural network approach to unsupervised learning. The algorithm operates through competitive learning, where neurons in the network compete to represent input data vectors. Each neuron has a weight vector of the same dimension as the input data, and during training, the network identifies the best matching unit (BMU) for each input vector.

The learning process involves two phases: competition and cooperation. During competition, the neuron with weights most similar to the input vector is selected as the BMU. In the cooperation phase, the BMU and its neighbors on the grid have their weights adjusted to become more similar to the input vector. The neighborhood function typically decreases over time, starting with a large neighborhood and gradually focusing on fine-tuning individual neurons.

SOMs can be configured with different grid topologies, including rectangular and hexagonal arrangements. The number of neurons and grid dimensions are key parameters that affect the level of detail in the resulting map. Training typically involves presenting the input data multiple times (epochs) until the network stabilizes and the weight vectors converge.

The resulting map provides a topology-preserving projection where similar data points are mapped to nearby locations, making it useful for clustering and visualization. The algorithm's ability to create ordered representations of complex data has made it a valuable tool in fields ranging from genomics to market analysis.

## Schema Markup
```json
{
  "@context": "https://schema.org",
  "@type": "Thing",
  "name": "self-organizing map",
  "description": "A machine learning technique for dimensionality reduction that maps high-dimensional data onto a lower-dimensional grid while preserving topological relationships",
  "sameAs": [
    "https://www.wikidata.org/wiki/Q173614",
    "https://en.wikipedia.org/wiki/Self-organizing_map"
  ],
  "additionalType": "data clustering algorithm"
}

## References

1. Freebase Data Dumps. 2013
2. [OpenAlex](https://docs.openalex.org/download-snapshot/snapshot-data-format)