6.2 Classification Rules 205 When producing rules using covering algorithms, missing values can be best treated as though they don't match any of the tests. This is particularly suitable when a decision list is being produced, because it encourages the learning algorithm to separate out positive instances using tests that are known to succeed. It has the effect either that instances with missing values are dealt with by rules involving other attributes that are not missing, or that any decisions about them are deferred until most of the other instances have been taken care of, at which time tests will probably emerge that involve other attributes. Covering algorithms for decision lists have a decided advantage over decision tree algorithms in this respect: Tricky examples can be left until late in the process, at which time they will appear less tricky because most of the other examples have already been classified and removed from the instance set. Numeric attributes can be dealt with in exactly the same way as they are dealt with for trees. For each numeric attribute, instances are sorted according to the attribute's value and, for each possible threshold, a binary less-than/greater-than test is considered and evaluated in exactly the same way that a binary attribute would be.