Preliminary theoretical results on a feature relevance determination method for Generative Topographic Mapping

摘要: Feature selection (FS) has long been studied in classification and regression problems, following diverse approaches resulting on a wide variety of methods, usually grouped as either /filters /or /wrappers/. In comparison, FS for unsupervised learning received far less attention. For many real problems concerning multivariate data clustering, becomes an issue paramount importance results have to meet interpretability actionability requirements. A method Gaussian mixture models was recently defined Law et al. (2004). Mixture are well established clustering but their visualization capabilities limited. The Generative Topographic Mapping (Bishop 1998a), constrained distributions, originally overcome such limitation. this brief report we provide the theoretical development feature relevance determination Mapping, based that (2004); with method, can be visualized low dimensional latent space interpreted terms reduced subset selected relevant features. [This documend revised (8/11/2006)]