| Summary: | 碩士 === 國立海洋大學 === 電機工程學系 === 92 === This thesis presents a novel approach to the autonomous solution for clustering, namely MMST+SAP which incorporates Modified Minimum Spanning Tree (MMST) and Synchronous Agglomerating Prototypes (SAP) to perform clustering with high accuracy. Fast clustering is made possible by involving the MMST pre-partition process followed by an agglomeration algorithm. Centroids of Mf subsets generated by MMST are taken as initial prototypes to be agglomerated into clusters by SAP. Each prototype is regarded as a neutral particle with an imaginary mass defined as the number of input vectors associated with the prototype. Each prototype Cj exerts an imaginary attraction force fij on all other prototypes Ci; the magnitude of fij depends on the distance between the two prototypes involved and the variance of data points associated with Ci. Due the attraction mechanism, in the course of SAP process each prototype will move toward the centroid of the cluster to which the prototype most likely belongs. The locations of prototypes are updated as all prototypes move synchronously in the feature space. After a certain steps of update, attraction forces between any pairs of prototypes would become so tiny that all prototypes virtually stop moving, and the SAP process is terminated. The prototypes moved to the same centroid as well as their associated data points are labeled as a cluster. Unlike conventional clustering methods such as k-means and the fuzzy c-means, MMST+SAP is free of the initialization problem and it needs not pre-specify the number of clusters, namely input partition is autonomously determined by the input nature. Extensive empirical results are provided to verify the performance of the proposed MMST+SAP clustering approach.
|
|---|
