CAKE – 合作式演化性知識擷取方法

Abstract

由於知識爆炸，知識可以分類為靜態知識和動態知識。儘管已有很多的知識擷取方法可以系統化的從領域專家那取得靜態知識的規則，卻沒有任何一種方法去深入討論發掘動態知識，其原因在於欠缺足夠的情境資訊(context information)。在本篇論文裡，我們提出一個方法命名為合作式演化性知識擷取(CAKE)的新知識擷取方法，其透過收集足夠的情境資訊來幫助專家發現動態物件的產生進而發掘動態知識。我們首先定義靜態個性化配置檔案(Profile)和動態行為來做為情境資訊並以服務敏感性和症狀相似性等事件來做為合作啟發(heuristic)以輔助專家意識到動態知識的發生。CAKE中我們用變異物件知識擷取(VODKA)和趨勢事件擷取(TEA)來建立動態擷取表格以增進發現變異物件的效率並自動的調整屬性順序表格(AOT)內物件與屬性之間的重要程度來更進一步發現演化性知識。這對協助專家了解屬性對物件之間的改變有很大的幫助。再來，我們設計CAKE讓動態EMCUD增加新的物件或新的物件屬性來更新已存在的知識表格並進而透過動態AOT將其原來的隱含規則具有適應的能力。除此之外，我們開發了一個電腦蠕蟲偵測雛型系統來驗證CAKE的效能。

Due to the knowledge explosion, the knowledge can be classified into static knowledge and dynamic knowledge. Although many knowledge acquisition methodologies have been proposed to systematically elicit rules of static knowledge from domain experts, none of these methods discusses the issue of discovering dynamic knowledge due to the lack of sufficient context information. In this thesis, we will propose a new collaborative knowledge acquisition methodology, Collaborative Acquisition for Knowledge Evolution (CAKE), to solve the issue of discovering dynamic knowledge by collecting sufficient relevant context information to help experts notice the occurrence of dynamic object. First, we define static profiles and dynamic behaviors as the context information to assist experts to be aware of the occurrence of dynamic knowledge based on several collaborative heuristics for service-sensitive and symptom-similar events. Variant Objects Discovering Knowledge Acquisition (VODKA) and Trend Event Acquisition (TEA) are used to construct a new dynamic acquisition table to facilitate the acquisition of variant knowledge and to automatically adjust the relative importance of each attribute to each object in the attribute ordering table (AOT) to discover the evolutional knowledge in CAKE. This is useful to help experts understand the changing behaviors of attributes to each object. Furthermore, CAKE is designed to use Dynamic EMCUD, a new version of an existing knowledge acquisition system called EMCUD which relies on the repertory grids knowledge acquisition technique to manage object/ attribute-values tables and to produce inferences rules from these tables, to update existing tables by adding new objects or new object attributes in new acquisition for adapting the original embedded rules with the dynamic AOT. Besides, a Worm detection prototype system is implemented to evaluate the effectiveness of CAKE.