Supervised classification with structured class definitions

Almost all work on supervised classification assumes that the class labels are a simple nominal variable. Often, however, there are relationships between the classes arising from an underlying structure. In particular, in prognostic situations, class memberships are often defined in terms of variables to be measured in the future. In such cases new forms of classification rules can be constructed, based on predicting these future variables instead of directly predicting the class memberships. We call such methods indirect methods. The circumstances in which one might expect indirect methods to outperform direct methods are described. Two simulation studies are outlined, the first showing that indirect methods can perform well in cases which are essentially impossible for conventional direct approaches, and the second illustrating some of the theoretical arguments. Two real examples, one on smoking and the other on retail banking, demonstrate that indirect methods can outperform conventional direct methods in real applications.