Model-Based Clustering with Nonparametric Maximum Likelihood

• Activity Number: 194
• Type: Contributed
• Date/Time: Monday, August 4, 2008 : 2:00 PM to 3:50 PM
• Sponsor: IMS
• Author(s): Yeojin Chung*+ and Bruce G. Lindsay and Jia Li
• Companies: The Pennsylvania State University and The Pennsylvania State University and The Pennsylvania State University
• Keywords: clustering ; nonparametric maximum likelihood ; EM algorithm ; HMAC ; MEM

Abstract

Clustering analysis is widely used as a major data mining technology. Li, Ray, and Lindsay (2007) developed the Modal EM(MEM) algorithm, searching for a mode of a mixture density starting from any given point. Applying MEM to the kernel density estimate, they derived the Hierarchical Mode Association Clustering (HMAC), grouping data points associated to the same mode. However, the kernel density estimate is naturally biased which could lower the performance of clustering. We investigate the improvement in HMAC by density estimators, based on treating the kernel density estimator as an element of the model consisting of all mixtures of the kernel, continuous or discrete. One can "likelihood tune" the kernel density estimator by using it as the starting value in an EM algorithm and it leads to a fitted density with higher likelihood and smaller bias than the kernel density estimator.