Evaluating learning models with transitions of human interests based on objective rule evaluation indices

This paper presents a method to support the evaluation procedure of a data mining process using human-system interaction. The post-processing of mined results is one of the key factors for successful data mining process. However, it is difficult for human experts to completely evaluate several thousands of rules from a large dataset containing noise. We have designed a method based on objective rule evaluation indices to support the rule evaluation procedure; the indices are calculated to evaluate each if-then rule mathematically. We have evaluated five representative learning algorithms to construct rule evaluation models of the actual data mining results from a chronic hepatitis data set. Further, we discuss the relationship between the transitions of the subjective criterion of a medical expert and the performances of the rule evaluation models.