Self-Organizing Map (Kohonen Map)
Also known as: SOM, Kohonen map, Kohonen network, öz-örgütlemeli harita
A self-organizing map is an unsupervised neural network, introduced by Teuvo Kohonen in 1982, that projects high-dimensional data onto a low-dimensional (usually two-dimensional) grid of prototype vectors while preserving the data's topology — nearby inputs map to nearby grid cells. It is used for visualization, clustering, and exploratory analysis, turning complex data into an ordered, interpretable map.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use a self-organizing map to visualize and explore high-dimensional data on an ordered 2D grid, to cluster while preserving neighbourhood structure, or to build interpretable maps of complex datasets — common in exploratory data analysis, process monitoring, finance, genomics, and document/sensor mapping. Its strength is the topology-preserving, human-readable layout (the 'U-matrix' reveals cluster boundaries). It requires choosing the grid size and shape, is sensitive to initialization and learning schedule, gives a non-deterministic result, and is more a heuristic visualization/quantization tool than a model with a likelihood. For pure 2D visualization, t-SNE and UMAP often separate clusters more sharply; for vector quantization/clustering, k-means is simpler — but neither gives SOM's ordered grid.
Strengths & limitations
- Produces an ordered, topology-preserving 2D map that is easy to visualize and interpret.
- Combines clustering and dimensionality reduction in one model.
- Robust, online-trainable, and scales to large datasets.
- The U-matrix and component planes give rich exploratory views.
- Requires choosing grid size/shape; results depend on these and on initialization.
- Non-deterministic and sensitive to the learning-rate/neighbourhood schedule.
- A heuristic quantizer, not a probabilistic model — no likelihood or principled selection.
- Can distort distances; the grid imposes a fixed topology that may not match the data.
Frequently asked
How does a SOM differ from k-means?
Both learn prototype vectors, but a SOM arranges its prototypes on a fixed grid and updates each winner together with its grid neighbours, producing an ordered, topology-preserving map. k-means has no inter-cluster geometry. So a SOM is k-means-like quantization plus a neighbourhood structure that yields an interpretable 2D layout.
Is a SOM the same as t-SNE/UMAP for visualization?
All three reduce dimensionality for visualization, but a SOM maps onto a fixed grid of prototypes (good for ordered, quantized maps and the U-matrix), while t-SNE/UMAP place individual points in a continuous plane and usually separate clusters more sharply. They serve different exploratory purposes.
What is the U-matrix?
The unified distance matrix visualizes the distance between each SOM node and its neighbours on the grid. High values form 'walls' between dissimilar regions, so the U-matrix reveals cluster boundaries on the trained map — the standard way to read clusters out of a SOM.
Sources
- Kohonen, T. (1982). Self-organized formation of topologically correct feature maps. Biological Cybernetics, 43(1), 59–69. DOI: 10.1007/BF00337288 ↗
- Kohonen, T. (1990). The self-organizing map. Proceedings of the IEEE, 78(9), 1464–1480. DOI: 10.1109/5.58325 ↗
How to cite this page
ScholarGate. (2026, June 2). Self-Organizing Map (Kohonen Map). ScholarGate. https://scholargate.app/en/machine-learning/self-organizing-map
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- K-Means ClusteringMachine learning↔ compare
- Locally Linear EmbeddingMachine learning↔ compare
- t-SNEMachine learning↔ compare