Does GSOM need labeled examples?

The basic learning approach is unsupervised; labels can still be used afterward to interpret or evaluate the map.

Does the map grow forever?

No. Growth follows specified rules and training conditions, rather than an obligation to add units indefinitely.

Is a growing map always better than a fixed one?

No. The better choice depends on the data, task, settings and evaluation criteria.

AI Glossary / AI and machine learning / Intermediate

Growing self-organizing map

Definition

A growing self-organizing map is an unsupervised neural-network method that expands its map of representative units as it learns the structure of data. It adapts map size instead of requiring every unit to be fixed in advance, as in a conventional fixed-grid self-organizing map.

What it means

A self-organizing map represents data using prototype vectors arranged in a lower-dimensional neighborhood structure. Similar observations tend to activate nearby units, making the map useful for exploring patterns.

Choosing a fixed map size can be awkward. Too few units may compress distinct patterns together, while too many can create unnecessary detail. A growing self-organizing map, commonly GSOM, addresses this by adding units during learning according to its growth mechanism.

Growth is driven by how well existing units represent the observations. The original GSOM work includes a spread factor to influence the amount of growth and therefore the granularity of the resulting representation.

The map does not automatically discover the only correct number of real-world categories. Its shape depends on feature choices, scaling, training settings and data coverage. A unit is a learned representative, not necessarily a meaningful business segment by itself.

How it works

Start with a small map whose units contain prototype vectors in the same feature space as the observations. Scale and encode features so that the distance measure reflects the intended notion of similarity.

Present observations and identify their best-matching units. Update the winning prototype and nearby prototypes so that they better represent the input. Neighborhood learning helps organize the map rather than treating prototypes as unrelated points.

Accumulate representation error and apply the growth rules when the specified threshold is exceeded. New units allow the map to represent regions that the existing structure describes inadequately. Boundary handling and initialization follow the particular algorithm.

After growth, refine the representation and inspect its quality. Evaluate quantization error, neighborhood behavior and the meaning of the patterns. If business categories are needed, an additional interpretation or grouping step may be required.

Three practical examples

1. A customer-analytics team explores patterns in product usage without labels. It investigates a growing map when it does not know in advance how much representational detail is useful.

2. An equipment researcher maps operating conditions from several measurements. Additional units may help distinguish densely populated or poorly represented regions of the observations.

3. A teaching exercise compares fixed and growing self-organizing maps on the same synthetic dataset. Students examine how scaling and growth settings affect the visual structure.

Think of it this way

Think of arranging samples on a display board and adding space where the current arrangement is too crowded to show useful differences. The board grows where its representation needs detail. The analogy is limited because GSOM uses numerical distances and explicit update rules; it does not know which distinctions people will find meaningful without those distinctions being reflected in the data and evaluation.

Worked example

Suppose a map has three prototype locations on one illustrative measurement: 2, 5 and 8. Observations at 2 and 5 are represented exactly, while an observation at 12 is four units from its nearest prototype, 8.

That large mismatch contributes more representation error than an observation at 8.5, whose nearest-prototype distance is only 0.5. A growth mechanism can respond to accumulated mismatch by allocating more representation capacity in a relevant region.

This example explains the motivation, not the full multidimensional GSOM rule or its exact growth threshold. A single outlier should not automatically be interpreted as a valuable new segment. The analyst must distinguish meaningful underrepresentation from bad data and unusual but irrelevant observations.

Illustrative case study

Fictional case: Birch Learning Services studies how organizations use its training catalog. The team has several usage measures but no agreed segmentation labels.

It trains a growing map, compares multiple runs and examines the organizations represented by different regions. Analysts notice that one strong separation is mostly driven by organization size, so they reconsider scaling before assigning business meaning to the picture.

The resulting map is used for exploratory discussion and follow-up analysis. Staff do not treat every unit as a marketing persona or assume neighboring units have identical needs. This case is illustrative and reports no real commercial outcome.

Common misconceptions and mistakes

Mistake: Assuming growth removes all design choices. The distance measure, features, learning settings and growth control still shape the result.

Mistake: Equating map units with final clusters. Prototypes organize observations, while useful categories may require another grouping or interpretation step.

Mistake: Trusting a two-dimensional layout as a perfect picture of all original distances. Dimensional reduction can distort relationships and should be checked against the underlying data.

Business use

GSOM can be useful for exploratory pattern discovery when a fixed map size is hard to justify. It may help teams discuss complex usage or operational data through representative regions.

Before using the result for decisions, assess stability, representation quality and whether segments remain useful on new observations. Keep a clear distinction between exploratory visualization and validated decision rules. The cost of interpreting and maintaining the map can matter as much as the training computation.

Check your understanding

A GSOM creates many units after one feature is measured on a much larger numerical scale than the others. What should be checked before interpreting the map?

Answer: Feature scaling and the resulting distance measure, because they may dominate both matching and growth.

Related terms

Unsupervised learning
Quantization error
Clustering
self-organizing map
prototype vector
feature scaling
topology preservation

Learn next

Learn a conventional self-organizing map first. Trace how one observation updates its winning prototype and neighbors, then examine how a representation-error threshold could motivate adding capacity.

Sources

Alahakoon, Halgamuge and Srinivasan, Dynamic Self-Organizing Maps with Controlled Growth for Knowledge Discovery: https://research.monash.edu/en/publications/dynamic-self-organizing-maps-with-controlled-growth-for-knowledge/. Institutional record consulted for attribution; numerical examples and fictional case are original. Editorial review on 14 October 2026 consulted the indexed primary abstract at https://pubmed.ncbi.nlm.nih.gov/18249788/ and author-uploaded paper excerpts at https://www.researchgate.net/publication/3302776_Dynamic_self-organizing_maps_with_controlled_growth_for_knowledge_discovery_IEEE_Transactions_on_Neural_Networks_113_601-614 for error accumulation, boundary growth, smoothing and spread-factor control.

Previous: Grover's algorithm Next: Growth curve (statistics)