Summary: | 博士 === 國立成功大學 === 資訊工程學系碩博士班 === 95 === The dissertation presents five novel clustering algorithms. The first one is the Adaptive C-Populations (ACP) clustering algorithm. The algorithm is capable of identifying dense regions, as well as influential minor prototypes, in an unlabeled dataset. The second is referred to as the Possibilitic Latent Variable (PLV) clustering algorithm. The algorithm replaces the binary latent variable with the fuzzy latent variable and rejects the use of the Gaussain probability density function in favor of a possibilitic function to measure the cost for associating point with clusters. The algorithm is suitable for many different kinds of populations, not solely for compact Gaussian distribution. The third is entitled as the Aggregative Fuzzy Latent Variable (AFLV) clustering algorithm. It extends the method of the PLV algorithm, and is designed to analyze the ordinal data. The algorithm evaluates membership degrees of objects to mixtures in each attribute according to accumulative occurrences of reference values, and aggregates the degrees of dependent mixtures from different attributes to construct final clusters. The fourth is the Fuzzy Cover Clustering (FCC) algorithm. The fuzzy cover in the algorithm is used to identify the holding points in the dataset. Since these holding points are always located in dense regions, they can service as the backbones of final clusters. The results evolve naturally to reflect actual groups in the data. The last algorithm, designated as the Variation-based Co-expression Detection (VCD) algorithm, is proposed to detect co-expression patterns from time-vary gene expression data. The algorithm adopts the cover and features a new measurement criterion for calculating the degree of change of the expressions between adjacent time points and evaluating their trend similarities. It is also unnecessary to determine the number of clusters in advance since the algorithm automatically detects those genes whose profiles are grouped together. The performances of the proposed algorithms are carefully verified by conducting clustering tasks on the contents of datasets with different characteristics.
|