| Summary: | 碩士 === 世新大學 === 資訊管理學研究所(含碩專班) === 95 === Classification is a widely used technique in data mining. It explores the association among attributes to contruct the classification rules, then uses them to classify instances or provide prediction for unseen data. Traditional rule mining only emphasizes single criterion, such as classification accuracy, or aggregates several criteria into a single evaluation function for the rule generation. Such results might not comply with the essence of multi-objective rule mining. In this paper, a well-known evolutionary multi-objective optimization algorithm, named Strength Pareto Evolutionary Algorithm (SPEA), is used to generate the non-dominated rules for the transaction data of an insurance broker company. The objectives used to evaluate rules are comprehensibility, accuracy, and interestingness. The rules generated from SPEA are compared with those from PRISM which is a rule induction algorithm. The results show that SPEA gets better scores in all objectives and coverage than PRISM. The best rule set obtained from SPEA is closer the Pareto-optimal front. Although the rules generated from SPEA is less than those of PRISM, it is much easier to for users to choose from some meaningful and interesting rules instead of choosing from numerous and mediocre rules. Finally, evolutionary multi-objective optimization algorithms are more suitable than traditional classification method for multi-objective rule mining.
|
|---|
