| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 94 === Unsupervised clustering is an important technique in pattern recognition, image processing, data analysis and data mining. There are many traditional nonhierarchical clustering methods have been used widely, but the most problems are that they need a priori information about the number of clusters and the best position of the initial centers. The fuzzy clustering has also been adopted in popular. In addition, the weighting exponent (fuzzifier) is another predefined variable which significantly affects the result of fuzzy clustering. In this paper, we proposed a two-stage unsupervised fuzzy and probabilistic clustering algorithm. In first stage, we use the concept of minimization of the error of the reconstructed dataset and the objective function, in order to decide the weighting exponent、the number of clusters and the position of the candidate centers. In second stage, we decide the final optimal clusters by probabilistic (EM) algorithm. According to the results for testing dataset, the accuracy of the proposed algorithm is higher than 94%, and is more stable and efficient than other traditional methods.
|
|---|
