Is a decision list the same as a decision tree?

No. A tree branches into different paths, while a list tests one ordered sequence until the first match. Some representations can be converted, but their compactness can differ.

Must outputs be only true or false?

Classical definitions often use Boolean outputs, but practical rule lists can return multiclass labels, probabilities or actions, with the exact semantics specified.

AI Glossary / Graphs and symbolic AI / Intermediate

Decision lists

Q: Can conditions overlap?

Yes. The first matching condition determines the output.

Definition

A decision list is an ordered sequence of rules in which the first matching condition determines the output. If no listed condition matches, a final default supplies the result. The order is part of the model's meaning.

What it means

A decision list resembles an “if, else if, else” procedure. Each rule pairs a condition with a prediction or action. Once a condition succeeds, later rules are not consulted for that input.

This makes a short list relatively easy to inspect: an explanation can identify the first satisfied rule and the earlier conditions that were not satisfied. It also means that reading each rule independently can be misleading, because a later rule applies only to cases that escaped every earlier one.

Lists can be written by people or learned from labeled examples. In machine learning, the goal may balance predictive error against list length or rule complexity. A compact list is interpretable only if its conditions and the data behind them are understandable.

How it works

Define the available input features and the conditions allowed in each rule. A condition might be one Boolean feature or a conjunction such as “temperature high AND inspection overdue.” Different decision-list families restrict these conditions differently.

At prediction time, evaluate conditions from top to bottom. Return the first matching rule's output, or the default if all fail. An overlapping pair of conditions is not inherently contradictory because precedence resolves which output is used.

Learning methods search for rule conditions, their order and their outputs. Greedy methods make sequential choices; optimization methods may search more broadly under a particular objective. A result described as optimal is optimal for that specified objective and candidate space, not necessarily for every business consideration.

Evaluate the resulting list on data not used to fit it. Examine how many cases reach each rule, the errors among those cases and behavior when features are missing or change distribution. Reordering rules is a model change and should be assessed as such.

Three practical examples

A support router first handles missing account information, then recognizes a product category, and finally sends unmatched requests to a general queue.
A classifier uses a short learned list of feature tests to produce a label with an inspectable decision path.
A quality-control workflow places mandatory review conditions before routine handling rules so that overlapping cases follow the intended precedence.

Think of it this way

Think of a reception desk using a priority checklist: the first applicable instruction decides where a visitor goes. A later instruction may also describe the visitor, but it does not override the earlier match.

The analogy assumes first-match semantics. Other rule systems may combine several matching rules, vote across them or resolve conflicts differently; those are not equivalent behaviors.

Worked example

Order changes the routing

Use this illustrative list:
1. If required details are missing, return “Request information.”
2. Else if the ticket is urgent, return “Priority review.”
3. Else return “Standard review.”

Suppose 100 tickets contain 10 with missing details, 25 marked urgent and 4 belonging to both groups. Rule 1 captures all 10 missing-detail tickets. Rule 2 then captures only the urgent tickets not already captured: 25-4=21. The default receives 100-10-21=69.

Now swap the first two rules. “Priority review” captures all 25 urgent tickets. “Request information” captures the 10-4=6 remaining missing-detail tickets. The default still receives 69.

The four overlapping tickets change output even though the conditions and their individual frequencies are unchanged. This is a routing illustration, not a recommendation that urgency or completeness must always take priority. The correct precedence depends on the workflow's requirements.

Illustrative case study

A fictional support-tool team, Maple Desk, learns a short rule list from previously labeled tickets. A reviewer moves a broad product rule above a missing-information rule because the product condition appears more common.

The reordered list sends some incomplete tickets directly to a specialist queue. The rules still look individually reasonable, but their combined behavior has changed. The team adds overlap examples to its review and reports how many tickets each rule captures after earlier rules have run.

It also inspects the default group rather than treating it as harmless leftovers. Unmatched cases can include new products or poorly represented requests, so the default requires an intentional output and ongoing evaluation.

Common misconceptions and mistakes

Rule order is merely presentation. It changes which matching condition wins.
A later rule describes all inputs satisfying its condition. It describes only those that also failed earlier conditions.
A short list cannot overfit. Specific conditions selected from many candidates can still fit accidental training patterns.
A default output means the system is confident. It may simply mean no other rule matched.

Business use

Decision lists can be useful where teams need a compact, inspectable prediction or routing mechanism. They support explanations that follow the actual computation rather than a separate approximation of it.

Assess whether first-match precedence reflects the intended policy. Track coverage and error for each reached rule, including the default, and evaluate the effect of missing features. When the list is learned, preserve its training objective, data version and evaluation evidence. When it is edited manually, treat the edited list as a new model requiring a focused check of changed cases.

Check your understanding

A ticket matches rules 2 and 5, and no earlier rule. Which output does a standard decision list return?

Answer: Rule 2's output. Rule 5 does not override the first match, even if its condition seems more specific.

Related terms

Decision tree
Default rules
Classification
Rule list
rule-based classifier
default rule
interpretable machine learning
model complexity

Learn next

Create three overlapping rules for an invented routing task and list one input matching each possible overlap. Evaluate them in two orders. Record the cases whose outputs change and decide which ordering matches the stated requirements.

Sources

Marchand and Sokolova, Learning with Decision Lists of Data-Dependent Features, defines first-match semantics and the default output: https://www.jmlr.org/papers/volume6/marchand05a/marchand05a.pdf

Angelino and colleagues, Learning Certifiably Optimal Rule Lists, examines learning under an explicit accuracy-and-complexity objective: https://www.jmlr.org/papers/volume18/17-716/17-716.pdf

The ticket counts and Maple Desk case are original illustrations.

Previous: Decision forest Next: Decision network