Predict labels or similarity matrices from a cluster fit

Convert posterior draws from a dpmixgpd_cluster_fit object into either a representative clustering or a posterior similarity matrix (PSM). This is the main post-processing step for the cluster workflow after dpmix.cluster() or dpmgpd.cluster().

Usage

# S3 method for class 'dpmixgpd_cluster_fit'
predict(
  object,
  newdata = NULL,
  type = c("label", "psm"),
  burnin = NULL,
  thin = NULL,
  return_scores = FALSE,
  psm_max_n = 2000L,
  ...
)

Arguments

object

A fitted cluster object.

newdata

Optional new data containing the response and predictors required by the original formula. New-data prediction is available only for type = "label".

type

Prediction target:

• "label": representative partition via Dahl's least-squares rule

• "psm": posterior similarity matrix on the training sample

burnin

Number of initial posterior draws to discard.

thin

Keep every thin-th posterior draw.

return_scores

Logical; if TRUE and type = "label", include the matrix of Dahl-cluster assignment scores.

psm_max_n

Maximum training sample size allowed for type = "psm".

...

Unused.

Value

A dpmixgpd_cluster_labels object when type = "label" or a dpmixgpd_cluster_psm object when type = "psm".

Details

Let \(z_i^{(s)}\) denote the latent cluster label for observation \(i\) at posterior draw \(s\). The posterior similarity matrix is $$ \mathrm{PSM}_{ij} = \Pr(z_i = z_j \mid y) \approx \frac{1}{S} \sum_{s=1}^S I(z_i^{(s)} = z_j^{(s)}). $$ The returned label solution is the Dahl representative partition, obtained by choosing the draw whose adjacency matrix is closest to the PSM in squared error.

For newdata, the function combines draw-specific component weights and component densities to produce posterior assignment scores relative to the representative training clusters. Returned newdata label objects also carry the training labels and response data needed for comparative plot(..., type = "summary") displays. A PSM is not defined for newdata, so type = "psm" is restricted to the training sample.

Computing the PSM is \(O(n^2)\) in the training sample size, so psm_max_n guards against accidental large matrix allocations.

See also

dpmix.cluster(), dpmgpd.cluster(), summary.dpmixgpd_cluster_fit(), plot.dpmixgpd_cluster_fit(), summary.dpmixgpd_cluster_labels(), summary.dpmixgpd_cluster_psm().

Other cluster workflow: dpmgpd.cluster(), dpmix.cluster(), plot.dpmixgpd_cluster_bundle(), plot.dpmixgpd_cluster_fit(), plot.dpmixgpd_cluster_labels(), plot.dpmixgpd_cluster_psm(), print.dpmixgpd_cluster_bundle(), print.dpmixgpd_cluster_fit(), print.dpmixgpd_cluster_labels(), print.dpmixgpd_cluster_psm(), summary.dpmixgpd_cluster_bundle(), summary.dpmixgpd_cluster_fit(), summary.dpmixgpd_cluster_labels(), summary.dpmixgpd_cluster_psm()