Can SIMCA reject all known classes?

Yes. A new observation can be incompatible with every modeled class.

Can it accept more than one class?

Yes. Separate class models can overlap, so multiple acceptance is possible.

Is SIMCA simply ordinary PCA followed by choosing the nearest class?

No. It builds class-specific models and applies compatibility criteria whose details depend on the variant.

AI Glossary / AI and machine learning / Intermediate

Soft independent modeling of class analogies

Definition

Soft independent modeling of class analogies, usually abbreviated SIMCA, is a class-modeling approach that builds a separate principal-component model for each class of interest. A new observation is assessed for compatibility with those models and may fit one class, several classes, or none.

What it means

Many classifiers force every input into one of the known categories. Class modeling asks a different question: does this observation resemble the range of examples that defines a particular class?

SIMCA represents variation within each class using principal components. These components describe a lower-dimensional structure in the class's measurements. Distances relative to that structure help assess whether a new observation is compatible with the class model.

The word soft refers to the possibility of overlapping or absent assignments rather than a compulsory single label. It does not automatically mean fuzzy membership probabilities. The word independent concerns separately constructed class models, not a claim that all measured variables are statistically independent.

SIMCA has several variants with different distance statistics and acceptance rules. The exact formulation matters when interpreting a result or comparing implementations. A generic reference to SIMCA is not enough to reproduce a classification decision.

How it works

Collect representative training observations for each class being modeled. Prepare the measurements through appropriate centering, scaling, and other justified preprocessing, keeping those choices consistent for future observations.

Fit a principal-component model to each class and select its complexity using suitable validation. Too few components can miss genuine within-class variation; too many can absorb noise or make the model overly permissive.

Measure how a new observation relates to each fitted model. Depending on the variant, assessment can consider its distance within the modeled component space and its residual distance away from that space.

Apply documented acceptance criteria separately to the class models. Evaluate accepted, rejected, and multiply accepted observations on independent data. The practical meaning of those outcomes depends on the task and on how well the training data represent legitimate class variation and relevant alternatives.

Three practical examples

1. A laboratory teaching exercise models spectra from a known material class and tests whether new spectra are compatible. A rejection signals a need for investigation rather than identifying a specific alternative material.

2. A quality-analysis team builds separate models for two product categories. It records cases accepted by both models instead of forcing an unsupported unique assignment.

3. A researcher changes the number of retained components and studies how acceptance behavior changes. This demonstrates the tradeoff between representing genuine variation and accepting unrelated observations.

Think of it this way

Think of creating a flexible outline around examples of each familiar shape. A new shape can fall inside one outline, inside overlapping outlines, or outside all of them.

The analogy captures separate class regions. Actual SIMCA models use multivariate geometry and statistical criteria, not hand-drawn outlines. Being inside a region means compatibility under the model; it does not prove an object's identity, origin, or suitability for every use.

Worked example

Suppose a simplified demonstration uses two normalized distances for each class: one within the component model and one residual distance. For this exercise only, both must be at most 1 for acceptance.

Observation X has distances (0.7,0.6) for class A and (1.3,0.4) for class B. It is accepted by A and rejected by B under this illustrative rule.

Observation Y has (0.8,0.9) for A and (0.6,0.7) for B, so both accept it. That is an ambiguous compatibility result, not necessarily a software error. Actual SIMCA variants use their own statistics and boundaries; these invented values illustrate independent decisions rather than a universal formula.

Illustrative case study

Fictional case: Cedar Materials Lab studies practice measurements from two reference material groups. It builds separate class models and reserves some samples for evaluation.

Several test samples are accepted by both models. The team examines whether the groups genuinely overlap in the chosen measurements or whether the models are too broad. Other samples are rejected by both, prompting checks of measurement quality and whether the known classes cover the input population.

The team reports these outcomes explicitly. This invented case illustrates class modeling and ambiguity. It does not claim that SIMCA alone certifies material authenticity or replaces domain-specific testing requirements.

Common misconceptions and mistakes

Do not assume soft means that the output is a calibrated probability. Do not confuse independent class models with independent measurement variables.

Do not interpret rejection as identification of a particular alternative class. An observation may be unusual because of a new class, a measurement problem, or an unrepresented legitimate condition.

Avoid fitting preprocessing or choosing component counts using the final test set. Do not report only overall accuracy when multiple acceptance and rejection are meaningful outcomes. Finally, name the SIMCA variant and its decision criteria.

Business use

SIMCA can be useful when compatibility with a known class matters more than forcing every observation into a category. Teams may encounter it in multivariate measurement analysis and quality-related research.

Evaluate false acceptance, false rejection, ambiguous acceptance, and coverage of legitimate variation. Define how each outcome enters the workflow, including when further examination is needed. Keep the model's statistical decision distinct from broader claims about quality, authenticity, or operational suitability that require additional evidence.

Check your understanding

A sample is rejected by class A. Does that prove it belongs to class B?

Answer: No. Rejection only indicates incompatibility with A under the chosen model and threshold. B must be assessed separately, and other explanations remain possible.

Related terms

Principal component analysis
Outlier detection
Class modelling
one-class classification
residual distance
multivariate analysis
acceptance region
chemometrics

Learn next

Learn how PCA models variation within one group. Then distinguish a discriminative classifier from a class model. Compare acceptance and rejection outcomes on a small dataset, documenting preprocessing, retained components, and the specific SIMCA decision rule used.

Sources

Svante Wold, Pattern Recognition by Means of Disjoint Principal Components Models: https://www.sciencedirect.com/science/article/pii/0031320376900145. Indexed primary abstract checked; publisher access still failed. mdatools authors, SIMCA classification: https://mdatools.com/docs/simca.html. Official documentation independently consulted for PCA class models, score/residual distances, and variant-dependent limits, resolving the generic-scope verification gap. The acceptance rule is explicitly illustrative, independently checked, and not a universal SIMCA formula; the case is fictional.

Previous: Soft clustering Next: Soft label