Rule based classifier; coverage and accuracy

A rule-based classifier classifies records using a collection of IF-THEN rules. A rule has the form

R: IF (condition) THEN (class)

where the condition (antecedent or precondition) is a conjunction of attribute tests and the consequent is the predicted class. Example: R1: IF age = youth AND student = yes THEN buys_computer = yes.

A rule covers a record if the record satisfies its condition; it fires when it is used to classify the record. Rules may be extracted from a decision tree or learnt directly (sequential covering).

Rule assessment measures

Let D be the training data, n_covers the number of records covered by rule R, and n_correct the number of those covered records that R classifies correctly.

• Coverage = n_covers / |D|. It is the fraction of all records the rule applies to.
• Accuracy = n_correct / n_covers. It is the fraction of covered records classified correctly.

Example (buys-computer data, |D| = 14): R1 covers records 9 and 11 (young, student = yes); both are "yes".

• Coverage(R1) = 2/14 = 14.3%
• Accuracy(R1) = 2/2 = 100%

A good rule has both high coverage and high accuracy. If several rules fire for a record, conflict resolution uses size ordering (the rule with most attribute tests), class-based ordering or rule ordering (decision list); if none fires, a default rule is used.