Accuracy of the Rules:

When automatically classifying new objects based on a rule set with given accuracies, the accuracies should not be taken literally. A 100% rule may not be correct for all objects in the problem domain, even though it was correct for all training examples. Still, the system doing classification will need to take these values as being correct, since no other information is available. The following four important cases could be discerned:
Only One Deterministic Rule Matches the New Object: If we have one rule which is said to be 100% correct, and this is the only rule that match the object, that rule's classification should be followed.
Only One Non-Deterministic Rule Matches the New Object: If only one rule match the object which shall be classified, but this rule is not 100% correct in the training set, its advice should be followed if the accuracy is above some given threshold, for instance 75%. Otherwise, the system may respond with no classification for this object, or try to find a rule which almost fit.
More Than One Rule Match New Object: In case more than one rule match the new object, several possibilities arise.

• Best rule The rule with highest accuracy could be used.
• Voting The different rules could give a vote to their classification, and one could sum up the total number of votes for each classification. This vote could be the accuracy or some other measure of the rule quality.
  

No Rules Match the New Object: If this happens the system might give no classification or try to find rules which almost match the object. How this could be done is described in [Ste93].