Kohonen Maps

Implementation of Self-Organizing Maps (SOM), Growing SOM (GSOM), and Neural Gas for unsupervised clustering and dimensionality reduction.

---

What's in this repo

Algorithm	Description	Notebook
SOM	Standard Self-Organizing Map on fixed grid	notebooks/01-som-basics.ipynb
SOM	Customer segmentation real-world use case	notebooks/02-som-customer-segmentation.ipynb
Neural Gas	Free neuron placement, no grid constraint	notebooks/03-neural-gas.ipynb
GSOM	Growing SOM — dynamic grid expansion	notebooks/gsom-iris-python.ipynb

---

Quickstart

Install

git clone https://github.com/abhinavralhan/kohonen-maps.git
cd kohonen-maps
python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install -r requirements.txt

Run a notebook

jupyter notebook

Navigate to notebooks/ and open any notebook.

Use the package directly

from kohonen.som import SOM
from kohonen.neural_gas import NeuralGas
from sklearn.datasets import load_iris

iris = load_iris()
X = iris.data

# Train SOM
som = SOM(grid_x=10, grid_y=10, learning_rate=0.5)
som.fit(X, epochs=1000)
som.plot_umatrix()
som.plot_heatmap(X, labels=iris.target_names[iris.target])

# Train Neural Gas
ng = NeuralGas(n_neurons=30)
ng.fit(X, epochs=2000)
ng.plot_neurons(X, labels=iris.target)
ng.plot_convergence()

---

How SOMs Work

SOMs, invented by Teuvo Kohonen in the 1980s, are neural networks that produce a low-dimensional (typically 2D) representation of high-dimensional data while preserving topological structure. Similar data points stay close on the map.

Algorithm

1. Initialize a grid of neurons with random weight vectors
2. For each input, find the Best Matching Unit (BMU) — neuron with weights closest to the input
3. Update the BMU and its neighbors to be more similar to the input
4. Decrease learning rate and neighborhood size over time
5. Repeat until convergence

Key Parameters

Parameter	Description
Grid Size	Dimensions of the neuron grid (e.g. 10×10)
Learning Rate	How much weights adjust per step — decreases over time
Neighborhood (sigma)	Radius of neurons affected by each update
Epochs	Number of training iterations

Visualizations

Plot	What it shows
U-Matrix	Distances between neurons — dark = cluster boundary
Hit Map	How many data points land on each neuron
Component Planes	How each input feature varies across the map

---

How Neural Gas Works

Neural Gas (Martinetz & Schulten, 1991) places neurons freely in data space — no grid constraints. Instead of grid distance, it uses rank of distance to determine neighborhood influence.

Key difference from SOM

	SOM	Neural Gas
Neuron placement	Fixed grid	Freely in data space
Neighborhood	Grid distance	Rank of distance
Best for	Visualization, topology mapping	Irregular clusters, density estimation
Interpretability	High (U-Matrix)	Medium

---

Choosing the Right Algorithm

Situation	Use
Need a 2D visual map of data	SOM
Clusters have irregular shapes	Neural Gas
Dataset grows over time	GSOM
Want topology + flexibility	GSOM
Benchmarking against PCA/t-SNE	SOM

---

Comparison with Other Methods

Method	Preserves Topology	Interpretable	Scalable
SOM	Yes	High	Medium
Neural Gas	Partial	Medium	Medium
GSOM	Yes	High	Medium
PCA	No	Medium	High
t-SNE	Partial	Low	Low
UMAP	Yes	Medium	High
K-Means	No	High	High

---

Repo Structure

kohonen-maps/
├── kohonen/                  # Reusable Python package
│   ├── __init__.py
│   ├── som.py                # SOM class
│   ├── neural_gas.py         # Neural Gas class
│   ├── gsom.py               # GSOM (coming soon)
│   ├── metrics.py            # Evaluation utilities (coming soon)
│   └── visualize.py          # Visualization helpers (coming soon)
├── notebooks/
│   ├── 01-som-basics.ipynb
│   ├── 02-som-customer-segmentation.ipynb
│   ├── 03-neural-gas.ipynb
│   └── gsom-iris-python.ipynb
├── requirements.txt
├── CONTRIBUTING.md
└── README.md

---

Applications

• Customer Segmentation — group customers by purchasing behavior
• Document Clustering — organize text by topic
• Anomaly Detection — identify outliers as points far from any neuron
• Image Compression — reduce color palettes while preserving visual quality
• Bioinformatics — gene expression analysis

---

References

• Self-Organizing Maps — Conceptual Overview
• GSOM Paper
• Neural Gas — Martinetz & Schulten (1991)
• MiniSom Library

---

Contributing

Feel free to get in touch!