應用貝氏網路於結構方程模式因果關係之研究

Application of Bayesian Networks to Infer Causality for Structural Equation Modeling

蘇招恭(Jaukong Su)

指導教授 ： 劉湘川

亞洲大學/資訊學院/資訊工程學系碩士班/碩士(2006年)

摘要

中文摘要
在統計學上，一般的統計分析，多偏重於兩變數間的相關性，而無法表示其變數間之因果關係，以身高體重為例，當身高愈重者，他的體重或許也愈重，兩變項間呈現正相關，然而在事實上，是身高影響體重？抑體重影響身高？若改變一個人的體重並不會影響其身高，所以體重並非是影響身高的原因。但在現實生活或社會科學的研究上，有許許多多的變數，其因果關係並非如上例那麼顯而易見，故在變項上，何者為因？何者為果？一直是學者很想探究的一個議題。
雖然貝氏網路近年來在各種研究上，藉由專家知識結構及學習模式，在因果關係及診斷上扮演著重要角色，亦被廣泛使用於醫學、資料探勘等領域之研究。
然而因果關係的推斷大都來自理論及長時間之觀察或實驗之資料，且貝氏網路目前並不能控制潛在變項，須藉由專家知識來建立，這對社會科學的研究似乎緩不濟急，故從資料中發現可能存在的潛在變項及因果關係便相對的重要。
本論文將在d隔離及淨相關的條件下，分析各個觀察變項並分群至同一潛在變項，以建立符合結構方程模式之測量模式，並結合PC2演算法及條件機率，仿建置貝氏網路之方法，建構以貝氏網路為基礎之結構方程模式。

關鍵字

貝氏網路 ； 結構方程模式 ； 科技接受模式

並列摘要

Abstract
In statistics, general statistical analysis stresses on the relevance between the variables. But it is unable to express the causality between the variables. Take the height and body weight as example. Height higher, his body weight is also heavier. It is connected between the variables. As a matter of fact, however, does the height affect the body weight or the body weight affects the height? If the change of the body weight would not affect the height, it is not the cause that the weight can be affected by the height. There are lots of the variables in real life or in the researches of social science. The causality is not as obvious as the example that we have mentioned. On variable, which one is a cause? Which one is a result? Therefore, it is always an issue that the researchers like to find out in the statistic.
By means of the constructions of specialists and the mode of learning, the Bayesian networks play the important roles in the causality and the diagnoses, and they are used generally in the researches of medicine and the data survey, too. However, the educts of the causality come from the theories, the observations of a long time and the data of the laboratory. In the meantime, the Bayesian networks cannot control the latent variables so far. It must be constructed by specialists’ knowledge. These results in the research are too late to the social sciences. Therefore, it is very important to find out the latent variables and the causality that could be existed possibly.
In order to build a measurement model that matches the structural equation model, these variables will be analyzed individual and observed then to be divide into the same latent variables under the d-separation and partial correlation condition. By combining PC2 algorithm and conditional probability and after the Bayesian Networks, the structural equation model will be built base on Bayesian Network.

並列關鍵字

bayesian network ； Structureal Equation Modeling ； Technology Acceptance Model

參考文獻

[3] 洪新原、梁定澎、張嘉銘(2005)。科技接受模式之彙總研究。資訊管理學報，2(4)，211-234。

[9] Bollen, K.A. (1989). Structural Equations with Latent Variables. John Wiley and Sons, New York, NY.

[12] Cooper, G. and E. Herskovits . (1992) . A Bayesian Method for the Induction of Probabilistic Networks from Data. Machine Learning, 9, 309-347.

[13] Davis, F.D. (1989). Perceived Usefulness, Perceived Ease of Use and User Acceptance of Information Technology. MIS Quarterly, 13, 3, 319-340.

[17] Lee, B., A. Barua, and A.B. Whinston. (1997). Discovery and representation of causal relationships in MIS Research: A methodological framework. MIS Quarterly, 21, 1, 109-136.