SzymonNowakowski / DMRnet

This is a development version of DMRnet — Delete or Merge Regressors Algorithms for Linear and Logistic Model Selection and High-Dimensional Data.
1 stars 0 forks source link

Add $\tau$ parameter to GLAMER #57

Closed SzymonNowakowski closed 1 year ago

SzymonNowakowski commented 1 year ago

GLAMER needs a $\tau$ parameter (a threshold for the minimal in-between $\beta$ distance required to merge levels)

SzymonNowakowski commented 1 year ago

$\tau$ is a threshold for the minimal in-between distance required to merge levels. This merging algorithm is detailed here:

This algorithm is equivalent to cutting a dendrogram of a single-linkage clustering at height $h=\tau$.

The heights in which changes to a resulting model occur are exactly the heights at which clusters get merged in a dendrogram (dendrogram$height variable). The algorithm in glamer v. 0.3.4 calculates the dendrograms for all factors involved, and then takes all the heights together, sorts them and creates models cutting the dendrograms at those heights in increasing order.

Consequently, glamer v. 0.3.4 could be considered a net version of the algorithm in the sense of $\tau$ parameter (in the same manner, as a net of $\lambda$ values is considered) if not for a clustering used: v. 0.3.4 uses complete-linkage as a default.

So in order to make it fully net version, in the sense of both $\tau$ and $\lambda$ values, the only change that is in order is a change to the default clustering.

Starting with version 0.3.4.9002, glamer will be using single-linkage clustering as a default. All other algorithms will be using complete-linkage clustering as a default, as before.