A Note on Bayesian Information Criterion in Model-Based Clustering

• 발행기관: 한국자료분석학회
• 저자명: 정광모(Kwang Mo Jeong)
• 간행물 정보: 『Journal of The Korean Data Analysis Society (JKDAS)』Vol.7 No.5, 1517~1529쪽, 전체 13쪽
• 주제분류: 자연과학 > 통계학
• 파일형태: PDF
• 발행일자: 2005.10.30

영문 초록

In clustering analysis determining the number of clusters is a very crucial issue but there is no fully general and satisfactory criterion for every clustering algorithm. The model-based clustering methodology, to which we mainly focus our attention has some advantages over the classical heuristic methods in the sense that the Bayesian information criterion(BIC) can be used to find an optimal number of clusters. The BIC is an approximation to the integrated likelihood for selecting a clustering model in Bayesian approach. The regularity conditions for an approximation is not generally satisfied for the finite mixture model, so we need to check the performance of the BIC in both of the cases with the Gaussian and the non-Gaussian distributions. The performance of BIC will be compared via Monte Carlo simulation.

목차

1. Introduction

2. Model-Based Clustering Methodology

3. Monte Carlo Study

4. Conclusion

References

키워드

model-based clustering BIC mixture model model selection