Self-organising map

Description

An unsupervised neural network that arranges records on a 2-D grid preserving similarity, producing a segment map and per-variable component maps to interpret with stakeholders.

A self-organising map is a grid of nodes, each with a weight vector. For each record the closest node wins, and the winner and its neighbours are moved towards the record; neighbourhood size and learning rate shrink during training. The result preserves the topology of the input: similar records land on nearby nodes. Nodes can be labelled with known classes (unused nodes get a null label) so the map also classifies. Its main strength is visualisation: a cluster map plus one component map per variable lets analysts and business teams read segments together, alongside absolute and relative views. Performance degrades as dimensionality rises, and region labels are the analyst's interpretation. In UX, SOMs offer a visual alternative to k-means for exploring segments in survey or behavioural data with stakeholders.

Attributes

Loss type
Siloed
Effort
Days

Steps

  1. Select and standardise input variables.
  2. Initialise a 2-D grid of nodes (for example 10 x 10 to 20 x 20) with weight vectors.
  3. For each record find the winning node and move it and its neighbours towards the record.
  4. Shrink neighbourhood and learning rate over training and repeat until the map is stable.
  5. Draw the cluster map and one component map per variable; label nodes with known classes if available.
  6. Label regions and interpret them with stakeholders in working sessions, checking robustness across random initialisations.

Sources

  • Handbook of Data Mining (Ye, 2003)
  • Organizational Culture and Firms' Internationalization, Innovativeness and Networking Behaviour: Hofstede Approach (Szymura-Tyc & Kucia, 2016)
  • Pattern Classification (Duda et al., 2001)

Also known as

  • SOM
  • Self-organising map for visual segmentation
  • Kohonen map
  • Kohonen feature map
  • Component maps

Categories

  • Analytics & Behaviour
  • Computational & Machine Learning