Guide To Kohonen Self Organizing Map (SOM)

A SOM is an unsupervised neural network algorithm created by Teuvo Kohonen in the 1980s. It maps high-dimensional input data onto a lower-dimensional (often 2D) grid of neurons while preserving topological (neighbourhood) relations among the data.

Training involves: (1) initialising weight vectors for each neuron randomly; (2) for each input vector, finding the best matching unit (BMU) by Euclidean distance; (3) updating the BMU and its neighbours’ weights to move closer to the input; (4) gradually reducing learning rate and neighbourhood radius until convergence.

Key components include: an input layer (with one node per feature), a lattice (grid) of output neurons each with a weight vector, a distance metric (e.g., Euclidean) to find the BMU, a neighbourhood function which decays over time, and a learning-rate schedule.

The BMU is the neuron whose weight vector is closest to the input vector according to the chosen distance metric. It “wins” each iteration and triggers updates to itself and its neighbours.

SOMs are used for clustering, dimensionality reduction, visualization of high-dimensional data, pattern recognition, anomaly detection, feature extraction, and in domains such as finance, geoscience, image analysis and more.

Unlike k-means, SOMs preserve topological relationships (neighbourhood structure) on a grid. Unlike PCA, which is linear, SOMs handle non-linear mapping and give more interpretable visual maps of data.

Advantages: intuitive 2D/3D visual maps of complex data, ability to capture and preserve topology, good for unsupervised learning and many features, useful for exploratory data analysis.

Drawbacks include: training can be slow for large datasets, results depend on parameters (grid size, learning rate, neighbourhood radius), handling categorical data is less direct, and there's no explicit generative model of the data.

Grid size depends on data complexity, sample size and computational constraints. A too-small grid underfits (coarse clustering); a too-large grid may overfit and slow convergence. Often experimentation with quantization error / topographic error helps.

Yes — through online or incremental learning, SOMs can update weights with new inputs over time without retraining from scratch. This makes them suitable for dynamic or changing data environments.

A Kohonen Self-Organizing Map (SOM) is an unsupervised artificial neural network used for clustering and dimensionality reduction. It maps high-dimensional data onto a lower-dimensional grid, making complex patterns easier to interpret. In SEO or competitive analysis, SOM can help organize campaign data, identify similar patterns, and highlight top competitors based on search terms.

The SOM algorithm begins with initializing node weights randomly. It then selects a random input vector and calculates the Euclidean distance between this vector and each node’s weight vector. The node with the smallest distance becomes the Best Matching Unit (BMU). The algorithm next determines the BMU’s neighbourhood radius, which defines which nearby nodes will have their weights updated based on the input vector.

During training, each node within the BMU’s neighbourhood updates its weight vectors by adjusting toward the input vector. The neighbourhood radius and learning influence decrease over time, allowing the map to refine itself gradually. This process repeats for all iterations until the network stabilizes. The final result is a structured, interpretable 2D grid that groups similar inputs, enabling clearer insights for applications like campaign optimization and SERP improvement.