| Summary: | 碩士 === 國立臺灣大學 === 工業工程學研究所 === 92 === This thesis presents a wrap-around Self-Organizing Map associated with an automatic classification mechanism for data clustering. The proposed data clustering method consists of two procedures. At first data are mapped onto topologically structured neural neurons, represented as either a traditional SOM or the proposed wrap-around SOM. Then, in the second stage, the topology of the structured neural neurons and associated characteristic vectors are used by an automation classification algorithm to divide the linked neurons into sub-graphs. These sub-graphs are then the results of data clustering. The classification algorithm uses 2-mean techniques to automatically find the threshold for link cutting between connected neurons. Three topology models are investigated and studied, including the original 2-D topology, a reduced spanning tree for the original topology, and a regenerated complete graph from all neurons. Several numerical examples are tested, including an example with data distributed as chained two rings. Results show that only the proposed wrap-around SOM associated with the 2-mean method based classification mechanism can produce a correct data clustering result. The main advantages of using the proposed method are that the number of groups of the clustered data is automatically determined and only wrap-around topology can deal with problems of mutual inclusive data distribution.
|
|---|
