Must a decision region be connected?

No. The worked example has two separate pieces for A.

Is its area the probability of that prediction?

Not generally; probability depends on the input distribution.

Can thresholds change regions without changing learned scores?

Yes. The conversion from scores to labels is part of the decision rule.

AI Glossary / Mathematics and optimization / Advanced

Decision region

Definition

A decision region is the set of inputs that a classifier assigns to a particular output label. If h is the classifier and a is a label, its region is Rₐ = {x: h(x) = a}. It belongs to the classifier's feature space. It describes the model's prediction rule, not necessarily the true class membership of the inputs that fall within it.

What it means

Imagine assigning a color to each possible feature vector according to the model's prediction. All positions with the same color form that label's decision region. The colored area can have several disconnected pieces and need not have a simple geometric shape.
The region includes possible inputs beyond the observed training examples. Those parts of feature space may be sparsely represented or entirely unobserved. Seeing that a model assigns a label there does not establish that its predictions there are reliable.

How it works

A deterministic classifier combines its learned parameters, feature transformations and decision rule to assign a label. For binary classification, a threshold on a score can determine two regions. For multiple classes, comparing class scores with a specified tie rule can determine several regions.
If exactly one label is assigned to every allowed input, the label regions are disjoint and together cover the input domain. Decision boundaries mark where different predictions come arbitrarily close. The precise allocation of a boundary point depends on the rule's inequalities or tie handling. Changing a threshold can change the regions without retraining the underlying scoring model.

Three practical examples

A message classifier assigns each text representation to a category. Its decision regions lie in the numerical feature space used by the model, rather than directly on a map of words that readers can easily visualize.
A quality-control rule assigns a part to a review category when measured properties meet conditions. Examining its region helps an engineer identify which combinations trigger that prediction, including combinations not seen in the training sample.
A decision tree splits feature space into leaf-specific pieces. Several leaves can predict the same class. The region for that class is their union, so one class need not correspond to one leaf or one connected area.

Think of it this way

Think of coloring a map by a delivery company's routing code. Two separate neighborhoods may share the same code even when an intervening neighborhood uses another one. The code's region is the collection of locations receiving it. Its size does not tell you how many customers live there, and its boundary does not tell you whether the assigned service is appropriate.

Worked example

Let the input domain be the square 0 ≤ x ≤ 10 and 0 ≤ y ≤ 10. Define a classifier that predicts A when x < 3, or when both x ≥ 7 and y < 4. It predicts B for every other point in the square.
Region A therefore contains a left strip extending across the full y range and a separate lower-right rectangle. Point (1, 1) receives A because its x coordinate is below three. Point (8, 1) also receives A because it satisfies the second condition.
Their midpoint is (4.5, 1). It satisfies neither condition and receives B. Thus region A is not convex: joining two of its points can pass through the other label's region. The gap between its two pieces also makes it disconnected.
The inequalities settle boundary cases. Point (3, 1) receives B because the first condition is strict. Point (7, 4) receives B because y is not below four. Point (8, 3) receives A.
The left strip has area 3 × 10 = 30 and the right rectangle has area 3 × 4 = 12. Their union has area 42 within a square of area 100. This means 42% of the square's geometric area receives A. It means 42% of cases only if their feature vectors are distributed uniformly over that square; an actual dataset can have a very different distribution.

Illustrative case study

In a fictional inspection project, analysts use the worked rule as a teaching model. They increase the right-hand y threshold from four to six while leaving the left strip unchanged. Region A's area grows from 42 to 30 + 3 × 6 = 48.
A manager interprets this as a six-percentage-point increase in the fraction of inspected items predicted A. The analyst explains that the new region adds the band with x ≥ 7 and 4 ≤ y < 6. The operational change depends on how many actual items fall in that band.
If no relevant items occupy it, no predictions change in the evaluated population. If many occupy it, the effect can be substantial. The team measures changed predictions and errors on appropriate labeled evaluation records. The geometric expansion explains where changes occur; the evaluation data determine their frequency and consequences.

Common misconceptions and mistakes

A decision region is not a cluster discovered from the data, although a clustering-based classifier can induce regions. Nor is it a confidence interval. Inputs sharing a predicted label can have different scores or levels of evidential support.
A two-feature plot may show a slice with other features held fixed, rather than the complete high-dimensional region. Its apparent shape can change when those fixed values change. Also, a class region can combine several tree leaves; counting leaves is not the same as counting predicted classes.

Business use

Use decision regions to explain the combinations of features associated with a prediction and to investigate threshold changes or model behavior near boundaries. Pair the geometry with the distribution of real inputs and labeled outcomes. Document feature transformations, permitted input ranges and tie handling. Simple and complex regions both require evaluation.

Check your understanding

Why does the midpoint receive B?

Answer: Its x coordinate is neither below three nor at least seven.

Why does (7, 4) receive B?

Answer: The right-hand condition requires y < 4.

What determines the operational effect of expanding a region?

Answer: The relevant inputs and outcomes in the newly included portion.

Related terms

Feature space
Decision boundary
Decision tree
Convex set
Classifier
decision threshold
class score
input distribution

Learn next

Compare regions produced by linear classifiers and decision trees. Then examine how feature transformations, thresholds and the input distribution affect the interpretation of a decision-boundary plot.

Decision boundary

Sources

Aalto Dictionary of Machine Learning, decision region, decision boundary and classifier: label-based subsets of feature space and their boundaries. https://aaltodictionaryofml.github.io/ADictML_Main.pdf The two-piece region, boundary checks, area calculations and fictional inspection case are original teaching illustrations.

Previous: Decision nodes Next: Decision support system