Robust Clustering for Longitudinal Data with Fast Computational Algorithm

• Main Authors: Jun-Kai Ke, 柯竣凱
• Other Authors: Hong-Dar Wu
• Format: Others
• Language: zh-TW
• Published: 2017
• Online Access: http://ndltd.ncl.edu.tw/handle/92183938012897774786

碩士 === 國立中興大學 === 統計學研究所 === 105 === This study introduces noncentrality and its variant as a measure for clustering longitudinal data,along with an improvement on the computational algorithm:the iterated recombination of subsets(IRS).Since the IRS algorithm still has considerable burden in computation,our improvement reduces the computation time from K to the nth power to Kn,where K is cluster number and n is the total sample size. The noncentrality parameter is computed under normality assumption.For longitudinal data,however,this assumption can be violated from time to time.As a remedy,we propose to amend the statistic that was used to estimate the noncentrality by replacing the original observations by the corresponding ranks.For comparisons,we also report the performance of a WSS-index,which was the ”within-cluster sum of squares” based on original observations.We apply our proposal to the clustering of two data sets: (i)Taiwan’s monitoring data of fine particulate matter(PM2.5),and (ii) the nuclear-magnetic-resonance(NMR) data of ”ginseng” with different ages.