運用貝氏定理下的模型選擇

Model selection in Bayesian method

張富琛(Fu-Chen Chang)

指導教授 ： 鄭明燕

國立臺灣大學/理學院/數學研究所/碩士(2010年)

https://doi.org/10.6342/NTU.2010.03042

摘要

本篇文章主要是在討探在多母體下的模型選擇。在這裡我們使用無母數的方法來做分析。我們想要找到一個判別準則有效率的辨別一群資料中任意一個資料點屬於哪一個母體。在這裡用的是貝氏的區分法則，建立在貝氏定理的基礎上。而且我們建立兩種估計的方法來決定兩種判別法則，也就是density estimation和logistic regression。現在的重點是找到一個方法能有效的增加這估計函數的準確度，而提升判別法則的精準度。在本文採用local likelihood density estimation和local logistic regression來估計未知函數，所以有兩種判別準則。藉由的帶寬選擇，以最小的AIC當做選取的標準能充份的對未知模型做選擇。在本文的模擬中顯示，只要樣本點數足夠多，此方法的確能有效改進選擇的成功率。

關鍵字

模型選擇 ； 貝氏區分法則 ； 局部最大可能性機率密度函數估計 ； 局部最大可能性邏輯迴歸分析 ； 帶寬選擇

並列摘要

This paper is concerned with the problem of model selection among multiple populations. Here we use a nonparametric approach. We would like to find a decision rule to effectively identify which population each data comes from. We create a decision rule, based on Bayesian theorem, called Bayesian discriminant rule. Furthermore, we construct two estimated methods to decide the decision rule – density estimation and logistic regression. An unknown density function has to be estimated in the decision rule. Now it is vital to find an accurate way to estimate this density function or logistic probability to arise the classification rate. By using bandwidth selection for local likelihood density estimation or local logistic regression that minimizes AIC criterion does improve the results of model selection. A small simulation shows that for a large enough sample size, the method performs well.

並列關鍵字

model selection ； Bayesian discriminant rule ； Local likelihood density estimation ； local logistic regression ； classification ； Bandwidth selection ； AIC

參考文獻

[1] C. Loader (1999), Local Regression and Likelihood: Springer.

[2] S. T. Chiu (1991), "Bandwidth selection for kernel density estimation," Ann. Statist, vol. 19, pp. 1883-1905.

[3] Tibshirani, R. J. and T. J. Hastie (1987). Local likelihood estimation. Journal of the American Statistical Association 82, 559-567.

[5] Pregibon, D. (1981). Logistic regression diagnostics. The Annals of Statictics 9, 705-724.

[6] Cook, R. D. (1977). Detection of influential observations in linear regression. Technometrics 19, 15-18.