本論文主要是運用自我映射組織圖(Self Organizing Map, SOM)可保留資料散佈結構的特性,將SOM與一些聚類(Clustering)演算法做一結合,成為一個改良後的兩階段聚類法,並透過一些範例加以證明本研究方法的可行性。 在本研究中,先後提出了兩個與SOM結合的聚類方法:其一,先以SOM對原始資料產生一初始聚類(Protoclusters)[21]或量化資訊(Quantization Information),再透過改良式的鄰近限制聚類法(Contiguity-Constrained Clustering Method)將第一階段產生的初始聚類,以總離散程度最小化(Minimum Global Variance)的方法,進行聚類合併的動作,以達到最後的聚類結果。 第二種方法則是將SOM與模糊最小最大聚類類神經網路(Fuzzy Min-Max Clustering Neural Network)[28]做一結合的聚類方法,主要將SOM作為一種前處理(Preprocess)的工具,將未排序過的資料順序做一排序動作,使得相似性高的資料得以被排序在一起,主要用來解決模糊最小最大聚類演算法(Fuzzy Min-Max Clustering Algorithm)會因為資料輸入順序的不同,而產生不同的超立方體(Hyperbox)分佈結果,進而造成聚類結果有所誤差的問題。 This thesis proposes two hybrid clustering approaches using neural networks. These clustering approaches use a Self-Organizing Map (SOM) to preprocess data, and apply some traditional clustering methods (e.g. Fuzzy Min-Max Neural Networks [28]) to data mining. Finally, we use some bench-mark exemplification for testing our approaches. The first approach uses the property of data topology preservation of SOM to generate protoclusters [21] (or quantization information) in the first step. Subsequently, quantization information will be clustered again by our improved Contiguity-Constrained Clustering Method that can obtain a minimum global variance when some closer protoclusters are merged. The second approach uses SOM to preprocess data and the result is applied to a Fuzzy Min-Max Clustering Neural Network (FMM) [28] for clustering. Such an approach can, therefore, be a solution for the sensitivity problem; that is, different input sequences of the same data set to Fuzzy Min-Max Clustering Algorithm may give different Hyperbox results.