類神經網路已經被成功地應用於解決各種分類及分群的問題,目前已有相當多的類神經網路模型應用在資料探勘領域。自1986 年起,類神經網路本身的運作一直被視為一個黑箱作業,缺乏對訓練結果的解釋能力,難以判斷類網路學習結果的合理性,因此提供一套合理及有效的類神經網路解釋方法是重要的。類神經網路適合應用於大量資料的資料探勘,提供資料的預測、分類、規則萃取、規則驗證、分群、自我組織及回歸等應用,從訓練後的類神經網路中可以萃取出有用的資訊及知識。法則萃取技術主要能為類神經網路提供解釋能力,使得類神經網路決策透明化及一般化。 本文提出一套類神經網路系統法則萃取的方法;利用自動分群的技巧從自我組織類神經網路中萃取分群的規則(rule),解釋決策過程及結果,進而歸納出類神經網路中的知識。本研究以三種資料集為例,經由上述的方法對資料集輸出的結果進行分群,並萃取出輸出結果的法則,以萃取出的法則應用於實際的資料分群,評估萃取出的法則資料分群的正確率,驗證此方法的可行性。實證結果顯示此方法所萃取出來的規則是合理的,所歸納出來有關資料集的知識也多數是合理的,本研究提供一套合理及有效的類神經網路法則萃取方法,挖掘規則來改善資料的一般性,以及從大量資料中建立法則來發現隱藏其中的知識,提供使用者驗證類神經網路的正確性,使得類神經網路的決策能夠透明化及一般化。
Neural networks have been successfully applied to solve classification and clustering problems. There are so many neural network models proposed in the field of data mining. Since 1986, the neural network has been considered a black box. However, it is short for the explanation of the training results. Thus, it is useful to provide a reasonable and efficient explanation method for the neural network. The neural network is suitable for dealing with a great number of data. It can extract the useful information and knowledge from the trained neural network. Rule extraction technologies mainly provide the explanation ability for the neural network and make the decision of the neural network transparent and understandable. In this thesis, we proposed a novel neural network rule extraction method. We use the auto clustering method to extract the clustering rules from the self-organizing feature map neural network and generalize the knowledge from its output map. We use three data sets to examine the feasibility of our proposed method. The experiment results also show that the rules extracted by our method are reasonable and the generated knowledge related to the data sets is also reasonable. Our proposed method can make the decision of the neural network transparent and understandable.