Identifying mixture models under model uncertainty

Beschreibung

The identification of finite mixture models when the number of components is unknown is considered. The first part of the talk sheds some light on
the role the prior of the weight distribution p(h) plays when the true number of components is unknown. It is shown that the very popular uniform
prior is usually a poor choice for overfitting models. A prior decision has to be made through the choice of p(h) whether for overfitting mixture
models empty components or identical, non-empty components should be introduced. As a consequence of this choice, either the number of nonempty
components or the total of components is a better estimator of the true number of components. In the second part of the talk identification
of finite mixture models that are strongly overfitting heterogeneity in the component-specific parameters is discussed. While standard priors lead
to underfitting the true number of components, shrinkage priors well-known from variable selection are applied to handle overfitting heterogeneity.
Such priors are able to discriminate between coefficients which are more or less homogenous and coefficients which are heterogeneous and avoid
underfitting of the number of components by reducing automatically the prior variance of homogeneous components.

Zeitraum	10 Dez. 2010 → 12 Dez. 2010
Ereignistitel	ERCIM’10, 3rd International Conference of the ERCIM WG on Computing and Statistics
Veranstaltungstyp	Keine Angaben
Bekanntheitsgrad	International