連續性屬性的離散化可以被視為如何去選擇出一組屬性切點集合的問題，多數的過去研究致力於找到一組最小的切點集合，並且同時保留資料的一致性。然而維持過高的資料一致性可能會導致分類演算法歸納出數目過多且概化能力不佳的分類規則。進行屬性離散化除了考量資料一致性外，也應該要將概化能力納入考量，因為概化能力好的分類規則是很容易被了解及解釋說明的。本研究中提出了以遺傳演算法為基礎的離散化方法，目標是能夠有效率地找出符合資料一致性及概化能力考量下的一個折衷最佳切點集合來進行離散化。本研究中設計了二組實驗，實驗中的資料選自於美國加洲大學爾灣分校的機器學習資料庫，實証結果顯示出本方法可以依照使用者的需求產生簡化的離散結果，而且可以幫助分類演算法歸納出概化能力佳及預測正確率亦高的分類規則。 Discretization of continuous attributes is one of main problems needed to be solved in data mining. Discretization can be viewed as the problem of selecting a set of cut points of attributes. Past studies concentrated on finding a minimal set of cut points and maintaining the fidelity of the original data in discretization. However, maintaining too high consistency may yield too many unnecessary rules which are not general. Generality is an important aspect to discretization because general rules are usually useful and easy to interpret. In this paper, a genetic algorithm based approach is proposed and the aim is to efficiently find an optimal compromise solution of discretization between generality and consistency criterions. Two sets of experiments on some data sets from UCI Machine Learning Repository by this approach were done. The empirical results have demonstrated that our GA approach can generate the simplest discretization result according to the requirement of the decision maker and help the classifier to induce general rules with high predictive accuracy.
