Decision Rules

A decision rule is an if-then statement: if debt ratio is above 50% and missed payments exceed one, route the application to manual review. Rules are compact enough to implement in a workflow, discuss with domain experts, or place directly into a policy document.

A decision list applies rules in order and stops at the first match. A decision set allows several rules to match the same case. Lists resolve conflicts through order; sets need an explicit conflict strategy.

Decision List vs. Decision Set

A decision list stops at the first matching rule. A decision set checks every rule and needs a conflict strategy when more than one fires. Select an applicant, then trace what each approach does.

Select applicant

Ama

income$78k
debt ratio22%
missed payments0
age under 30no

Rules

#1income ≥ 60k AND missed = 0approve
#2debt ratio ≥ 50%deny
#3missed payments ≥ 2deny
#4age < 30 AND debt ratio 30–50%review
defaultno rule matchedreview

Step through the list

#1income ≥ 60k AND missed = 0
Rule fires → stops here →approve
#2debt ratio ≥ 50%
#3missed payments ≥ 2
#4age < 30 AND debt ratio 30–50%
Lists are easy to audit: each case follows exactly one path. Sets are more expressive but require an explicit policy for conflicts — silent overlap produces unpredictable results.

Two numbers matter together:

  • Coverage or support: the share of cases where the rule applies.
  • Precision: among matched cases, the share with the predicted outcome.

A perfect rule that applies to two people may be less useful than a 90% accurate rule that covers half the population. The right balance depends on what happens after the rule fires.

Coverage Versus Precision

A lender needs a short rule for routing applications to human review. Compare a broad rule with a narrow one and watch what each rule catches and misses.

Matched cases

4 / 8

Coverage

50%

Precision

75%

#1

D 18%

M 0

paid

#2

D 48%

M 2

default

#3

D 54%

M 0

default

#4

D 31%

M 3

default

#5

D 42%

M 1

paid

#6

D 39%

M 0

paid

#7

D 62%

M 2

default

#8

D 27%

M 0

paid

A high-precision rule can still be operationally weak when it covers very few cases. Always report who the rule applies to alongside how often it is right.

Domain shapes the format

Financial services often favors ordered rules because staff can apply them consistently and document which condition triggered an action. A clinical team may prefer a small tree when the order of questions mirrors an existing diagnostic process.

Plan for overlap and silence

What happens when three rules match? What happens when none match? A deployed rule system needs conflict resolution and a default action. Leaving those cases implicit creates unpredictable policy.

Algorithms for learning rules

  • OneR — builds a classifier from a single feature; maximum simplicity.
  • Sequential Covering — learns rules one at a time, removing covered examples after each.
  • RIPPER — Sequential Covering with aggressive pruning to reduce overfitting.
  • SLIPPER — boosted variant of RIPPER; combines weak rule learners.
  • CN2 — Sequential Covering using a beam search to find high-quality rules.
  • Bayesian Rule Lists — selects an ordered list from pre-mined rules using a Bayesian prior over list length and rule complexity.
  • RuleFit — generates rules from a tree ensemble, then fits a sparse linear model over those rules and the original features.

OneR

OneR is the simplest rule-learning algorithm: it builds a classifier from a single feature. For each feature, it generates rules by reading off the most frequent class at each value. It then picks the one feature whose rules produce the lowest total error across the training set.

The algorithm runs in five steps:

  1. Discretize continuous features into intervals.
  2. Build a cross-table counting how often each feature value coincides with each class.
  3. For each feature value, write a rule predicting the majority class in that cell.
  4. Sum the errors across all rules for that feature.
  5. Select the feature with the smallest total error.

The result is a decision list with one condition per rule. OneR is interpretable by construction — a single feature drives every decision — but that same constraint limits accuracy when the outcome depends on combinations of features.

OneR in a screening context

A hospital triage tool that routes patients by a single vital sign is following OneR's logic. Staff can apply it instantly without a chart, and the rule can be audited against any dataset. The cost is that patients with an unusual combination of normal vitals and high risk may be misclassified.

OneR Algorithm Walkthrough

OneR picks the single best feature and turns each of its value ranges into a rule. Step through the algorithm to see how it scores and selects that feature.

Step 1 of 5

Discretize features

Feature

Continuous features are split into discrete bins. The split points come from the data distribution and domain knowledge — finer bins allow more expressive rules but risk overfitting.

Credit score < 580

9 instances

Credit score 580–669

8 instances

Credit score 670–739

8 instances

Credit score ≥ 740

10 instances

OneR is a useful baseline — if a richer model barely outperforms a one-feature rule, interpretability may win. Its weakness is that it cannot capture interactions between features.

Sequential Covering

Sequential Covering is a general procedure for building a decision list one rule at a time. Rather than learning all rules simultaneously, it peels off a layer of the data with each iteration:

  1. Start with an empty rule list.
  2. Learn the best rule on the current dataset.
  3. Add that rule to the list and remove all data points it covers.
  4. Repeat on the remaining data until no rule meets a quality threshold or all positive examples have been covered.
  5. Return the decision list.

Each rule is learned on a different subset, so later rules can focus on patterns that earlier rules missed. The final list is applied in order — the first matching rule fires. RIPPER and CN2 are concrete algorithms built on this procedure, differing mainly in how they measure rule quality and prune overfitting.

Order matters downstream

Because each rule is trained on whatever data the previous rules left behind, the list is order-dependent. Swapping two rules can change predictions even if both rules look correct in isolation. Document the training order alongside the rules themselves.

Sequential Covering Walkthrough

Sequential Covering learns rules one at a time. After each rule is found, the instances it covers are removed so the next rule can focus on what remains.

Start: full dataset

We have two classes mixed in feature space. Sequential Covering will learn one rule at a time, peeling off the data it explains.

x1x2

Rule list so far

No rules learned yet. The algorithm starts from scratch.
circletriangle
Each rule in the final list was trained on a different slice of data. That makes the list order-dependent — swapping rules can change predictions for edge cases near the boundaries.
Checkpoint

Rule A is 100% precise and covers 2% of cases. Rule B is 88% precise and covers 60%. Which is better?

Checkpoint

Design a default action for a loan-review rule set when no rule matches. What risk does your default create?