How to solve multi-class problems with DMS?
In a case when the target attribute has more than two classes (but not too many of them) it is possible to transform the problem to a series of two-class problems which can be solved by DMS. The transformation can be done in few different ways. Suppose that C is the number of classes in the original problem.
- A rule is constructed for every class from the original problem so that it is declared as the positive class and all other classes as the negative class. The result is an unordered set of C rules which may be used independently of other rules in the set
- A list of rules is constructed in a similar way as in the previous item but so that covered examples are eliminated from the data file. The result is a set of (typically) C-1 rules which must be used in the order they have been constructed.
- A decision tree of rules is constructed so that both positive and negative classes of the generated sub-problem are one or more classes of the original problem. The process is repeated until all generated sub-problems can be solved as two-class problems. The final solution has always less than C rules, the induced decision tree can reflect the structure of the domain but the process is not well defined because there are many different ways in which the original classes can be separated in subgroups.
In this system there is no way for automatic transformation of multi-class to two-class problems. But the system enables that only minor input data file modifications allow rule induction for different two-class sub-problems.
A good practice is to give class names in the original problem so that they are different from all other attribute names and values. When a two-class problem should be prepared then only the target class name (or names) has to be converted to a string beginning by character '!'. Additionally, if some class has to be completely eliminated from the induction process then its class name is substituted by a string which begins with an '?'.
© 2001 LIS - Rudjer Boskovic Institute
Last modified: January 23 2018 18:24:58.