本研究提出一套以階層式方法與粒子群演算法為基礎之ID3模糊決策樹,稱之為「階層式粒子群模糊決策樹(Hierarchical Particle Swarm Fuzzy Decision Tree; HPS-FDT) 」演算法。而設計一個高效能的模糊分類決策樹,必須考量歸屬函數(Membership Function)調整與規則庫(Rule Base)的設計,因為複雜的歸屬函數形狀需花費大量計算時間,而過多或過少的規則將會造成使用上的不易或使精確度降低。本研究利用粒子群最佳化調整歸屬函數形狀,並且為了有效控制規則數量,使用階層式方法控制屬性、語意個數,使其規則減少至適合的數量。本研究以三個主要目標為研究主軸,分別為「提高分類正確率」、「減少分類規則數量」、「降低使用到的資料屬性個數與語意個數」。另外以IRIS與WINE兩組資料進行實驗分析後,研究發現HPS-FDT不論在模糊規則數量或資料分類準確度皆有不錯的表現。 為了說明本研究提出的HPS-FDT可以運用在實際資料中,以網路銀行之基金交易資料作為實際案例;首先,利用顧客行為變數的RFM指標模式將顧客區隔分類。獲得每個顧客分類之後,將顧客基本資料以及顧客交易資料作為HPS-FDT欲分析之顧客特徵屬性,最後分析最佳化模糊決策樹所產生的顧客規則,瞭解顧客特徵屬性與顧客分類之間的關係,讓銀行業者能藉此針對不同價值之顧客做出適當的行銷策略。
Decision tree is one of most common techniques for classification problems in data mining. Recently, fuzzy set theory has been applied to decision tree construction to improve its performance. However, how to design flexile fuzzy membership functions for each attribute and how to reduce the total number of rules and improve the classification interpretability are two major concerns. To solve the problems, this research proposes a hieratical particle swarm optimization to develop a fuzzy decision tree algorithm (HPS-FDT). In this proposed HPS-FDT algorithm, all particles are encoded using a hieratical approach to improve the efficiency of solution search. The developed HPS-FDT builds a decision tree to achieve: (1) Maximize the classification accuracy, (2) Minimize the number of rules and (3) Minimize the number of attributes and membership functions. Through a serious of benchmark data validation, the proposed HPS-FDT algorithm shows the high performance for several classification problems. In addition, the proposed HPS-FDT algorithm is tested using a mutual fund dataset provided by an internet bank to show the real world implementation possiblility. With the results, managers can make a better marketing strategy for specific target customers.