Solution based on ILLM confirmation rule
Dragan Gamberger · 2001
The submitted solution was induced by ILLM (Inductive Learning by Logic Minimization) system. It is an Occam‘s razor based inductive system which presents knowledge in form of rules. The main building blocks for the rules are literals representing simple logical conditions of domain attributes. This knowledge representation form enables easy and comprehensible interpretation of induced results what was very important for the Coil challenge. The ILLM system can, as one of its options, generate rules in the so called confirmation rule set form. This form is typically used for applications requiring reliable target class prediction. The form consists of a set of confirmation rules so that every confirmation rule must be a simple conjunction of literals, every rule must be true for examples of the target class only, and every confirmation rule must be true for at least predefined number of target class examples defined by the selectable support level. Number of induced rules in the set is determined by the support level so that all confirmation rules satisfying this level are included into the set. Typically the user determines to include only few confirmation rules into the final set. In this case selection among all acceptable confirmation rules is based on their covering properties for the target class examples. In case of Coil challenge, confirmation rule set form is selected because such rule form is especially easy for human interpretation. This requirement restricted also the total number of induced rules in the set. The final solution consisted of only one confirmation rule. The main problem of confirmation rule induction in this domain was very high level of noise among training data. Typically the problem of noise in ILLM is solved in preprocessing by an explicit noise detection algorithm. In this approach, detected noise is eliminated from the training set before the rule induction process. In the Coil domain, additionally, it was necessary to allow that some non-target examples are covered by the induced rules as well. By selecting the optimal ration of the number of covered target class examples and the number of covered non-target class examples it was possible to induce rules covering about 20% of all examples, what was the condition in this competition.