| Summary: | 博士 === 中原大學 === 應用數學研究所 === 106 === Abstract
Spectral clustering is a clustering method that utilizes graph construction to represent
data with their similarities. Data points are represented by nodes, and similiarities be-
tween nodes are represented as weights of edges between nodes. The clustering process
can be viewed as a graph partitioning problem where in the clustering results, similar data
are connected. One of the challenges when applying spectral cluster for data analysis is
the construction of similarity function that is used to build the similarity matrix between
points. Another challenge is how to obtain the cluster number which is a difficult prob-
lem in the clustering method. This dissertation explores these challenges and the spectral
clustering application for cell formation problem in manufacturing systems. The Powered
Gaussian-kernel spectral clustering is proposed to build similarity function without any
user-defined parameters and furthermore, it can help to estimate the cluster number, si-
multaneously. For cell formation problem, a novel algorithm called direct k-way spectral
clustering is proposed. The algorithm has better performance and improve the recursive
2-way MinMaxCut algorithm for cell formation. Finally, minimum dissimilarity distance
had been added to the direct k-way spectral clustering in order to get an optimal number of
cells for machine-part cell formation. Experimental results and comparisons demonstrate
the superiority and effectiveness of these proposed methods.
Keywords: Spectral clustering, powered Gaussian-kernel, cluster number, cell formation,
minimum dissimilarities, cell number.
|
|---|
