Can a tree split on the same feature more than once?

Yes. Later nodes can introduce other thresholds to define finer regions.

Does a value exactly at the split go left or right?

That depends on the stated comparison. In the example's less-than-or-equal convention, it goes left.

Is the threshold always the average of all values?

No. A common candidate is the midpoint between adjacent distinct sorted values, which is a different calculation.

AI Glossary / Graphs and symbolic AI / Intermediate

Split point

Definition

A split point is a threshold used to divide examples according to the value of a numerical feature, especially in a decision tree. A test such as length less than or equal to 12 sends an example to one branch, while values above 12 go to the other. The threshold and selected feature together define the split.

What it means

A numerical feature can take many different values, so a tree needs a way to turn those values into branching decisions. A split point creates two groups at a particular node. Later nodes can apply further thresholds, including additional tests on the same feature.
The threshold is learned from the examples reaching that node, not necessarily from the complete training dataset. It is also different from a classification probability threshold applied to a model's final score. One partitions the feature space inside the tree; the other decides how an output score becomes an action or class label.

How it works

During training, a tree compares candidate features and thresholds using a split-quality criterion. For classification, that can be the weighted impurity of the resulting groups. For regression, it can be a measure of variation in the numerical targets.
A common exact-search approach considers thresholds between adjacent distinct observed feature values. It selects a candidate that improves the objective while respecting constraints such as minimum group size. Equality handling must be explicit: in a less-than-or-equal test, a value exactly equal to the threshold follows the left branch. Missing values require their own documented handling rather than an assumed numeric comparison.

Three practical examples

A classification tree can split labeled parcels by a measured dimension when learning a routing rule.
A regression tree can split observations by a numerical operating condition to predict different average outcomes in the resulting regions.
An analyst reviewing a tree can examine whether a learned threshold is supported by enough examples near its boundary.

Think of it this way

Think of placing a divider along a sorted row of cards. Cards on one side go into one group and the rest into another. The analogy is limited because a learned split is chosen using target information and an objective; the neatness of the division on the table does not establish that it predicts future cases well.

Worked example

Consider an original classification example with one numerical feature x. The six training values are 2, 4, 6, 8, 10, and 12. Their labels in that order are A, A, B, B, B, and A. There are three examples of each class, so the starting Gini impurity is 1 minus 0.5 squared minus 0.5 squared = 0.5.
Try a split at 5, using x less than or equal to 5 on the left. The left group contains two A examples and has impurity zero. The right group contains three B examples and one A example, so its impurity is 1 minus 0.75 squared minus 0.25 squared = 0.375. The weighted child impurity is (2/6) times 0 plus (4/6) times 0.375 = 0.25. The impurity reduction is 0.25.
Now try a split at 7. The left group contains two A and one B; the right contains one A and two B. Each group's impurity is 4/9, so the weighted result is also 4/9, about 0.4444. Its reduction is only about 0.0556. The split at 5 is better under this criterion.
Checking the other midpoints gives weighted impurities 0.4 at 3, 0.5 at 9, and 0.4 at 11. Therefore 5 is the best of these five candidates. Any threshold from 4 inclusive to 6 exclusive produces the same training partition under the stated comparison, but different thresholds can route a new value such as 5.5 differently. Equal training partitions do not mean identical behavior on all future inputs.

Illustrative case study

In a fictional sorting-project review, a tree contains a threshold of 5 on a measured feature. The team initially treats that number as if it were a physical boundary discovered with certainty. Looking at the training records, they find that no observed value lies between 4 and 6.
They document that the chosen midpoint separates the observed groups but is not a precise natural law. They also inspect measurement variation and new cases near the boundary. This scenario illustrates interpretation of a learned rule, not a claim about a real sorting system's performance.

Common misconceptions and mistakes

A split point is not necessarily one of the observed training values. A midpoint can represent a partition between them. Nor must the best local split produce the globally best complete tree; ordinary greedy growth makes local decisions.
Do not confuse a feature threshold with a train-test split or a final prediction cutoff. Do not ignore units: a threshold expressed in centimeters cannot be applied unchanged to values in millimeters. A seemingly precise decimal threshold can also reflect numerical computation rather than equally precise knowledge about the domain.

Business use

Split points make tree decisions easy to inspect, but they should be reviewed in context. Ask how many examples reached the node, what target criterion selected the threshold, and what happens to values near it.
When a tree supports an operational decision, evaluate the full model on suitable held-out data. A large reduction in training impurity at one node is useful evidence about that training partition; it is not by itself evidence of a causal boundary or guaranteed future accuracy.

Check your understanding

A node tests x less than or equal to 5. Where do values 5 and 5.1 go?

Answer: The value 5 goes left, while 5.1 goes right. That boundary convention must be preserved when implementing or explaining the rule.

Related terms

Learn next

Next, repeat the worked partition using threshold 5.8. All six training examples follow the same branches as at 5, but a new example with x = 5.5 switches from right to left. Explain why choosing a representative threshold is part of defining the predictive rule, not just a formatting choice.

Sources

AIMA original glossary sentences, split point test: https://raw.githubusercontent.com/aimacode/aima-glossary/master/sentences.txt scikit-learn, Decision Trees, mathematical formulation and classification criteria: https://scikit-learn.org/stable/modules/tree.html

Previous: Split-half reliability Next: Splitter