A New Effective Approach for Solving the Rules Conflict Problem

In this paper a new approach for the rules conflict problem is presented. The problem is discussed related to use of inductive learning systems generated rules. The rules conflict problem occurs when two or more rules cover the same test example but predict different classes. The common strategy for preferring one of the conflicting rules is the best rule strategy: for each rule a "weight" is calculated by some rule quality measure and the conflicting rule with the highest "weight" is chosen. In case of existing "weight" functions only training examples are used for calculation. In this paper we introduce a new rule "weight" function, called rule reliability, which is calculated on the basis of testing examples. It can reorder rule set so that rules with better classification "force" will be used firstly. Effectiveness of the new approach is shown by experiments.