The study of dynamic learning vector quantization in weighting adjustment

瀏覽統計 Email 通知 RSS Feed

簡歷

基本資料

Project title

Code/計畫編號

MOST104-2410-H019-035-MY2

Translated Name/計畫中文名

動態學習向量量化網路於權重值校正之研究

Project Coordinator/計畫主持人

Liang-Ting Tsai

Funding Organization/主管機關

National Science and Technology Council

Department/Unit

The Center of Excellence for the Ocean,National Taiwan Ocean University

Website

https://www.grb.gov.tw/search/planDetail?id=11727603

Year

2016

Start date/計畫起

01-12-2016

Expected Completion/計畫迄

31-07-2017

Bugetid/研究經費

447千元

ResearchField/研究領域

教育學

本研究計畫將以數值模擬方式，配合臺灣社會變遷調查（Taiwan Social Change）的實徵資料分析，驗證所發展的DLVQ (dynamic learning vector quantization)權重校正法，於多個背景變項的調查研究中，相對於多變數反覆加權法（Raking）及CRRE（classical raking ratio estimator）校正法，有較佳的穩定性及母群體推論正確性。資料分析的加權程序對於推論母群體特徵的正確性及重要性，可以從一些相關的研究中得到證實（例如：蔡良庭、楊志堅， 2008；楊志堅、蔡良庭，2008；蔡良庭，楊志堅，2014；Asparouhov, 2005, 2006； Kaplan & Ferguson, 1999）。而權重遺失時，申請者先前的研究所提出的LVQ 校正法（例如：蔡良庭，楊志堅，2014；楊志堅、蔡良庭、楊志強、2009；Shih, Tsai, & Yang, 2013；Tsai & Yang, 2012），則提供本研究計畫有良好的研究基礎。 Raking 與CRRE 都是利用邊際母體總數（marginal population）進行權重值校正，並無考慮每一個群體的原始樣本數與取樣數之間的關係，因此，當取樣不均勻，甚至因遺漏值的產生而導致分群取樣數出現0 時，將有可能產生權重校正錯誤（蔡良庭，楊志堅，2014；Lohr, 2010）。本計畫架構在原有的LVQ 校正法，提出DLVQ 法，除了利用邊際母體總數進行校正之外，更考慮受訪者的實際作答反應，以進行權重的計算。研究結果將有助於調查研究的權重校正，使分析的推論更正確。本計畫並呈現先導性實驗（pilot studies）的結果，這些資料顯示出本計畫的高度可行性及貢獻。This project is based on numerical simulation and empirical data analysis of Taiwan Social Change. We verified the developed DLVQ (dynamic learning vector quantization) for weighting adjustment of this project has better stability and effectiveness than the traditional Raking and CRRE (classical raking ratio estimator). Weighting process of data analysis plays an important role to infer the accuracy of population (Asparouhov, 2005, 2006; Tsai & Yang, 2008; Tsai & Yang, 2014). In the applicant’s previous studies about LVQ weighting adjustment (Tsai & Yang 2013; Tsai, Yang, & Yang, 2009; Shih, Tsai, & Yang, 2013; Tsai & Yang, 2012) can provide good foundations for this project. Raking and CRRE weighting adjustment are calculated by marginal population so that the relationship between each group’s original sample size and sampling size cannot be considered. Therefore, due to the inconsistent ratio of sampling size and original size or there are any cells with a zero value, the error of weighting adjustment will occur (Tsai & Yang 2014, Lohr, 2010). The proposed DLVQ method in this project not only calculates by marginal population, but also considers the actual responses of interviewees. The research findings contribute to the weighting adjustments and the analysis inference can be more precise. This project also presents the pilot study and relevant research results. These results indicate the high feasibility of current proposal and contribution of this research.

Keyword(s)

動態學習向量量化網路
多變數反覆加權法
CRRE
權重值校正
DLVQ
Raking
CRRE
Weighting adjustment

DSpace CRIS

The study of dynamic learning vector quantization in weighting adjustment

基本資料

Description