模糊知識探勘之多屬性決策

Abstract

在企業組織所使用的資料庫中隱藏著相當豐富的資訊；經由資料探勘技術，許多有用的規則將會被發掘出來。另一方面，柔性計算的使用在過去幾年擴展的非常的快速，而模糊集合則在其中扮演著關鍵性的角色。本論文進一步地使用模糊集合產生模糊知識的表示式。使用模糊集合的原因，在於以自然語言描述的模糊知識相當符合人們的主觀思考，也因此有助於增加使用者在決策時的彈性。此外，在設計一個模糊系統時需要考慮的一個準則，就是模糊的表達方式是否為使用者所理解；而簡單模糊分割是一個較被偏好使用的方法。 基於簡單模糊分割，本論文的目的在於發展獨特的模糊資料探勘方法，由多屬性資料中找出可理解與潛在性有用的模糊知識，並進一步結合柔性計算技術以應用於解決不同的決策問題。首先，針對關聯規則與序列樣式這兩個相當重要的知識表達方式，分別提出模糊關聯規則與模糊序列樣式的探勘方法；而模糊關聯規則是特別強調應用於多屬性分類問題的可行性。接著，使用由資料中找出的高頻樣式，例如複合技能與消費行為等，分別提出處理能力集合擴展與評估產品屬性權重等重要多屬性決策問題的方法。而獲取複合技能的目的，係由於某些有用的複合技能在加入能力集後，有助於獲取所有的單一技能，因此嘗試利用模糊資料探勘方法由單一技能中找出潛在有用的複合技能。 對於分類問題，係使用適應函數考量較少分類規則與較高正確分類率的基因演算法，以由訓練樣本中自動找出「若…，則…」型式的模糊分類規則。就分類一般化能力測試而言，在與其他模糊或非模糊分類方法的比較上，由使用 Iris data 與Appendicitis data 所得到之實驗結果，可看出所提出之方法擁有良好的分類能力。 在能力集合擴展方面，則有兩個議題被加以探討並提出可能的解決方法。首先是將由資料庫中所找出的模糊知識，視為組織決策者所需要的能力集，並使用 Feng 與 Yu 於 1998 年所提出之最小擴展表格法由能力集中產生具有最小學習成本的學習序列。其次則是以灰關聯度評估任意兩個單一技能間的學習成本，並進一步使用多層神經網路獲取由複合技能至單一技能的學習成本，而其目的在於避免以時間或金錢評估學習成本的不可能性。 至於在產品屬性的權重評估上，係針對不同的高頻率消費行為，分別評估消費者對產品屬性的重視程度。利用由交易資料庫中所找出的高頻率消費行為，以及由問卷所獲得之對產品與其各個屬性之偏好值，即可將每一個產品轉換為一單層感知器的訓練樣本。單層感知器經過訓練後，即可由其連接權重獲得消費者在不同的高頻率消費行為下，對產品屬性的重視程度。 經由數值範例或模擬結果，可說明與闡釋本論文所提出之各個方法可以有效的使用模糊知識，並提供有用的資訊，以支援多屬性問題之決策分析。此外，將灰關聯與自我組織特徵映射結合，且命名為「灰色自我組織特徵映射」的新分群技術也被提出。由使用 Iris data 所得到之實驗最佳結果，可看出灰色自我組織特徵映射優於其他非監督學習形式之神經網路。而灰色自我組織特徵映射也能夠有效解決旅行銷售員問題。

Most organizations have large databases that contain a wealth of potentially accessible information. Through data mining techniques, many interesting patterns or useful rules hidden in data will be discovered. On the other hand, soft computing techniques have expanded enormously over the past few years. Fuzzy sets are one critical component of soft computing, and are further used to generate fuzzy knowledge representations in this dissertation. The reason is that we consider that fuzzy knowledge representations described by the natural language are well suited for the subject thinking of human subjects and will help to increase the flexibility for users in making decisions. Additionally, the comprehensibility of fuzzy representation by human users is also a criterion in designing a fuzzy system. The simple fuzzy partition methods are thus preferable. The main aim of this dissertation is to develop novel fuzzy data mining techniques to find comprehensible and potentially useful fuzzy knowledge based on the simple fuzzy partition method; then those fuzzy knowledge, including fuzzy association rules, fuzzy sequential patterns and frequent patterns, are further applied to solve various multi-attribute decision problems by using soft computing tools. The feasibility of using fuzzy association rules in multi-attribute classification problems is specially explored. Subsequently, novel methods are further proposed by soft computing techniques to cope with two significant multi-attribute decision problems that include competence set expansion and assessment of weights of product attributes in individual purchase behaviors. Since some compound skills can be added to the needed competence set for helping to acquire all single skills, potentially useful compound skills are extracted from single skills. For classification problems, we employ genetic algorithms to automatically find fuzzy if-then rules from training patterns. In addition, the acquisition of a compact fuzzy rule set with high classification accuracy rate is taken into account in the fitness function. For classification generalization ability, the simulation results from the iris data and the appendicitis data demonstrate that proposed learning algorithm performs well in comparison with other fuzzy or non-fuzzy classification methods. For competence set expansion, two issues with possible solutions are discussed. First, the fuzzy knowledge can be treated as a needed competence set that should be acquired by decision makers; then, that needed competence set with minimum learning cost is expanded by the minimum spanning table method proposed by Feng and Yu (1998). Next, since it seems that it is not easy to measure learning costs by time or money, the other method is to obtain learning costs between any two single skills by using the grey relational grade. The learning cost from one compound skill to another single skill is further obtained by using a trained multi-layer neural network. As for the assessment of individual weights of product attributes, the focus is to assess weights or degrees of consumers’ attentiveness of product attributes for various frequent purchase behaviors. By using frequent purchase behaviors discovered from transaction databases, and evaluations of product attributes through questionnaire, each product can be transformed into a piece of input training data for a single-layer perceptron (SLP). After training SLP, the weights of products’ attributes in each frequent purchase behavior can be found from connection weights of SLP. Through numerical examples or simulation results, we illuminate that individual proposed methods can effectively use fuzzy knowledge to provide useful information to support multiple attributes decision making. Additionally, a new clustering technique, named the grey self-organizing feature maps (GSOFM), is proposed by incorporating the grey relations into the well-known self-organizing feature maps. From the simulation results, we can see that the best result of the GSOFM applied for analysis of the iris data outperforms those of other known unsupervised neural network models. Furthermore, the GSOFM can effectively solve the traveling salesman problems.