|
Abstract:
|
In many modeling applications, one model fits some observations well, while a different model fits other observations. The models that fit data in different regimes can be used to form clusters within the data. We view this situation similarly to that in which K means clustering is appropriate. While in the case of K means, the models in the clusters are simply means, we extend the clustering approach to more general models. Our objective is two-fold; we wish to identify clusters in the data, but we also wish to explore relationships of the variables to each other in each of the clusters. We investigate various approaches to this problem. All are computationally intensive. One method begins with groups based on K-means clustering, fits the best model in each from a catalog of models, and then iteratively moves outlying observations from one group to another and refits the models.
|
