Does it need known groups?

Yes, supervised discriminant analysis uses group labels.

Can it produce more axes than the number of features?

No; both feature count and group count constrain the usual discriminant space.

Is normality always perfectly satisfied?

Usually not. Assumptions and practical predictive performance should both be examined.

AI Glossary / AI and machine learning / Intermediate

Multiple discriminant analysis

Definition

Multiple discriminant analysis uses several measured variables to distinguish among known groups, commonly by constructing linear combinations that emphasize separation between groups relative to variation within them. The term is often used for multiclass linear discriminant analysis.

What it means

Suppose a dataset contains measurements from three labeled product types. The measurements may overlap individually, yet a combination of them may separate the groups more clearly. Discriminant analysis seeks such combinations using the known labels.

This makes it a supervised method. It differs from principal component analysis, which seeks directions of overall variation without using class membership. A direction with little total variation may still help distinguish groups.

The terminology varies across fields, so an analysis should state its exact model. In the common linear discriminant framework, group means and within-group covariance structure determine discriminant directions and classification rules. It is not a general name for every method involving several explanatory variables.

How it works

Training data provide measurements and group labels. The method estimates how group centers differ and how observations vary within groups. It then finds combinations that make between-group differences prominent relative to within-group spread.

For K groups and p measured features, there are at most min(p, K minus 1) discriminant dimensions in the usual formulation. A related probabilistic classifier assumes Gaussian class distributions with a shared covariance matrix and incorporates class priors.

The assumptions and covariance estimates need checking. With many features and few observations, regularization or dimension reduction may be necessary. Held-out classification performance matters more than separation seen on the training plot.

Three practical examples

A quality-control research exercise distinguishes labeled material types using several physical measurements.
A document-analysis team compares linear discriminant features with other representations for known document categories.
A researcher visualizes labeled groups in a compact supervised space, while separately checking how well new observations are classified.

Think of it this way

Think of adjusting a camera angle so that several overlapping groups become easier to distinguish in the picture. The chosen view uses knowledge of which objects belong to which group.

The analogy is limited: the method combines numerical features, not physical viewpoints. A clear training view can also exaggerate separation if the dataset is small or if information from test examples influenced the fitted directions.

Worked example

Why the number of groups limits the display

Imagine a study with three labeled groups and ten numerical measurements per observation. Standard multiclass linear discriminant analysis can provide at most two discriminant dimensions because three group centers have at most two independent directions of separation.

The analyst can therefore display the observations on two discriminant axes, even though the original records contain ten features. This is a supervised display: changing the group labels would generally change the axes.

Suppose the training plot appears well separated. That is not enough to establish useful prediction. The analyst applies the learned transformation and classification rule to observations held out before fitting, then records a confusion matrix for the three groups.

If one group is frequently mistaken for another, the attractive training plot has not translated into adequate generalization. The analyst should investigate sample size, assumptions and whether the groups genuinely overlap.

Illustrative case study

Fictional case: classifying material samples

Oakmere Materials, an invented laboratory team, collects several measurements from labeled samples of three material types. It tests a discriminant model as an interpretable baseline.

The first evaluation randomly splits measurements, but some measurements come from the same physical specimen. The team corrects the split so that specimens, rather than individual readings, are held out. Performance is then assessed on genuinely separate samples.

The case illustrates evaluation design, not a reported laboratory result. It does not establish that the method can certify material identity or replace the laboratory's required validation procedures.

Common misconceptions and mistakes

Multiple discriminant analysis is not multiple linear regression: the target is group membership, not an unrestricted continuous value. It is also not unsupervised clustering, because known labels guide the analysis.

Linear discriminant analysis and quadratic discriminant analysis make different covariance assumptions. Strong training separation does not prove that those assumptions hold, that classes are causally meaningful or that future observations will be classified accurately.

Business use

Businesses can use discriminant analysis as a baseline for labeled classification problems and as a way to inspect which combinations of measurements separate known categories. Its relative simplicity can support communication with domain specialists.

Check data quality, class balance and the number of independent observations per group. Compare linear, regularized and other suitable alternatives under the same validation design. Report class-specific errors rather than accuracy alone. When the cost of one confusion is much larger than another, incorporate those consequences into the decision policy. Treat coefficient interpretation cautiously when features are strongly correlated.

Check your understanding

There are four groups and twenty features. How many standard linear discriminant dimensions can be obtained at most?

Answer: Three, because the maximum is the smaller of twenty and four minus one. Fewer may be meaningful if the data have additional rank limitations.

Related terms

Learn next

Learn next why shared covariance produces linear decision boundaries while separate class covariances can produce quadratic boundaries. Then compare supervised discriminant plots with unsupervised PCA plots from the same labeled dataset.

Covariance

Sources

Primary implementation documentation: scikit-learn, https://scikit-learn.org/stable/modules/lda_qda.html ; estimator reference, https://scikit-learn.org/stable/modules/generated/sklearn.discriminant_analysis.LinearDiscriminantAnalysis.html . The scikit-learn guide sections on supervised reduction, shared covariance, class priors and the K-minus-one dimensional bound were consulted. Dimension example and fictional case are original.

Previous: Multiple correspondence analysis Next: Multiple inheritance