Reduced error specialization based on the information content of rule set

Except for over-fitting, excessive generalization should lead to high error rate of the learnt rule set, which is seldom discussed by literatures. When excessive generalization is occurred, the rule set will give multiple classification for a particular instance. The errors caused by generalization actually result in the increased inner conflict of the generalized rule set. In this paper, the inner conflict of rule set is defined based on the expanded knowledge of rules and a novel algorithm named RES(reduced error specialization) is proposed for the error rate reduction of rule sets. The best merit of RES is that it can eliminate the inner conflict of a rule set completely while the unknown knowledge of the rule set is unchanged. This fact will guarantee the error rate of the rule set on every test data will be determinedly reduced.