6.2 Classification Rules 203 trees than C4.5's pruning method. This has been investigated empirically by Oates and Jensen (1997). In our description of decision trees, we have assumed that only one attribute is used to split the data into subsets at each node of the tree. However, it is possible to allow tests that involve several attributes at a time. For example, with numeric attributes each test can be on a linear combination of attribute values. Then the final tree consists of a hierarchy of linear models of the kind we described in Section 4.6, and the splits are no longer restricted to being axis-parallel. Trees with tests involv- ing more than one attribute are called multivariate decision trees, in contrast to the simple univariate trees that we normally use. The CART system has the option of generating multivariate tests. They are often more accurate and smaller than univari- ate trees but take much longer to generate and are also more difficult to interpret. We briefly mention one way of generating them in the Principal Components Analy- sis section in Section 7.3. 6.2 CLASSIFICATIONRULES We call the basic covering algorithm for generating rules that was described in