Predictive Tests 1 Learning with Globally Predictive

We introduce a new bias for rule learning systems. The bias only allows a rule learner to create a rule that predicts class membership if each test of the rule in isolation is predictive of that class. Although the primary motivation for the bias is to improve the understandability of rules, we show that it also improves the accuracy of learned models on a number of problems. We also introduce a related preference bias that allows creating rules that violate this restriction if they are statistically signiicantly better than alternative rules without such violations. A variety of rule learning systems have been developed that create rules to predict class membership of examples such as AQ15 1], CN2 2], ITRULE 3], C4.5-ruless4], FOIL 5], FOCL 6], Greedy3 7], Ripper 8], and decision lists 9]. One commonly reported advantage of modeling predictive relationships with rules is the comprehensibility of the learned knowledge. Rule learners produce a set of learned rules of the form: