Learning with globally predictive tests

We introduce a new bias for rule learning systems. The bias only allows a r ule 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 mo tivation 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 proble ms. We also introduce a related preference bias that allows creating rules that violate this restriction if they are statistically significantly bette r than alternative rules without such violations.