Is GMDH always a polynomial neural network?

No. That is a prominent implementation, while the broader approach includes other forms and search strategies.

Does it use backpropagation in its classic polynomial form?

Candidate coefficients are commonly fitted through regression-style estimation and structural selection; this differs from ordinary end-to-end backpropagation.

Does external validation guarantee the selected model is best?

No. Search space, sample size, repeated selection, and future data shifts limit that conclusion.

AI Glossary / Deep learning / Intermediate

Group method of data handling

Definition

The group method of data handling, or GMDH, is a family of inductive modeling methods that generate candidate models, estimate their parameters from data, and select useful structures using an external evaluation criterion. A well-known form builds multilayer polynomial networks from small candidate functions. GMDH is associated with Alexey Ivakhnenko's work on automatically identifying model structure and complexity, rather than specifying every connection and layer in advance.

What it means

A modeling team often knows which variables might matter but not how complex their relationship should be. GMDH searches among progressively constructed descriptions and uses separate evaluation data or another specified external criterion to judge candidates. The objective is to balance fit with useful predictive structure. Despite its name, it is not a generic method for grouping files or a synonym for clustering. Polynomial networks are an important version, but the wider GMDH family includes other candidate forms and search procedures.

How it works

In a common multilayer iterative version, pairs of input variables form candidate units. Each unit may fit a quadratic polynomial containing a constant, both inputs, their product, and their squares. Coefficients can be estimated on a fitting subset because the polynomial is linear in those coefficients. Candidate outputs are evaluated separately, selected outputs become inputs to a later layer, and a stopping criterion limits further growth. The exact split, retained-candidate count, polynomial family, and stopping rule depend on the algorithm. Search restrictions mean that automatic construction should not be described as an unrestricted guarantee of finding the best possible model.

Three practical examples

Process approximation: construct candidate relationships between measured operating conditions and a continuous output.
Forecasting research: combine lagged measurements into polynomial predictors while respecting chronological evaluation.
Model-structure exploration: compare small candidate expressions when the team wants to inspect selected interactions and assess whether added complexity improves prediction.

Think of it this way

Think of assembling candidate explanations from small building blocks, testing them on examples they were not fitted to, and keeping promising combinations for another round. The analogy is limited because the selected explanation is a mathematical predictor, not necessarily a causal account, and repeated selection can itself adapt to the evaluation sample.

Worked example

Suppose a candidate unit has the form z = a0 + a1 x1 + a2 x2 + a3 x1 x2 + a4 x1 squared + a5 x2 squared. With coefficients [1, 2, -1, 0.5, 0, 0] and inputs x1 = 2, x2 = 3, its output is 1 plus 4 minus 3 plus 3 = 5. The coefficients are illustrative; fitting six coefficients requires sufficient independent information, not merely writing down the six terms.
Now compare two fitted candidates on three separate validation targets [4, 6, 8]. Candidate A predicts [5, 5, 8], giving squared errors [1, 1, 0] and mean squared error 2 divided by 3, about 0.667. Candidate B predicts [4, 7, 10], giving squared errors [0, 1, 4] and mean squared error 5 divided by 3, about 1.667. Under this chosen criterion, A ranks better, even if B fit the coefficient-estimation subset more closely. A later-layer candidate would need its own evaluation. These tiny constructed samples illustrate selection, not evidence that A will generalize reliably.

Illustrative case study

A fictional production team generates many polynomial candidates for predicting a quality measurement. It selects the network with the lowest error on a frequently reused validation subset. Before publishing the result, the team evaluates the selected pipeline on a separate later period and finds that some apparent improvement does not persist. It revises the search budget and evaluation procedure. The case illustrates that external selection criteria help guide complexity but do not make repeated model search immune to overfitting.

Common misconceptions and mistakes

A polynomial expression is not automatically simple once many layers are composed. Expanding it can create numerous high-degree interactions, making interpretation difficult. Variables selected by a predictive search are not necessarily causes. A low external error on a small sample can be unstable, and random splitting can leak future information in forecasting tasks. Automatic model growth also does not remove the need for preprocessing, candidate restrictions, domain checks, and comparison with simpler baselines.

Business use

GMDH may be useful for structured prediction problems where a controlled search over compact functions is attractive. Define the data available at prediction time, the candidate family, and the evaluation objective before comparing methods. Report the final selected structure and the full selection procedure, not just the winning score. For extrapolation, inspect how polynomial terms behave outside the observed range, because modest coefficients can still produce implausible outputs at extreme inputs. Operational value depends on reliable validation and maintainable complexity, not the historical label alone.

Check your understanding

Why should candidate coefficients and candidate structure not be judged only on the same training error?

Answer: More flexible candidates can fit that sample better without predicting new observations better. An external selection criterion helps assess that tradeoff, with a further independent evaluation needed for final reporting.

Related terms

Model selection
Validation set
Feature interactions
Overfitting
Polynomial neural network
regression
structural identification
feature interaction
inductive modeling

Learn next

fit two small polynomial candidates on one subset and rank them on another, then evaluate the selected model on untouched data. Study how multilayer composition, candidate pruning, and time-aware splits change the search and its reliability.

Sources

Ivakhnenko, Polynomial Theory of Complex Systems: https://www.gmdh.net/articles/history/polynomial.pdf GMDH implementation reference, candidate and criterion options: https://bauman-team.github.io/GMDH/python/html/gmdh.html GMDH algorithm descriptions: https://gmdh.net/GMDH_alg.htm

Previous: Group fairness Next: Group testing