|
Abstract:
|
Clustering methods assign observations to clusters based on the measured characteristics of each observation. In high-dimensional space, the true structure of clustering sometimes is contained in a low-dimensional subspace, the remaining dimensions contain little or no information about the clusters. Clustering methods that use all variables are often confounded by statistical noise in the dimensions that are unrelated to the cluster model. It is shown that a dimension-reduction technique, known as projection-pursuit, can increase the performance of the model-based clustering methods. A model-based projection-pursuit clustering methodology is introduced, which combines model-based clustering with projection-pursuit. In addition, the model-based projection-pursuit clustering provides low-dimensional pictorial representations of the clustering contained in the high-dimensional data.
|
