reRank: Serialize "1 vs. all" Feature Selection
In exprso: Rapid Deployment of Machine Learning Algorithms

Description Usage Arguments Details Value

This experimental function converts multiple feature rank lists, derived from "1 vs. all" binary feature selection, into a single feature rank list. This function is not in use in this package.

1	reRank(fss)

fss

The result of a doMulti function call.

After passing a feature selection method through doMulti, a set of ranked features gets returned for each one of the total number of levels in the $defineCase factor. In order to proceed with model deployment (at least in the setting of a conventional pipeline where feature selection occurs prior to classifier construction), multiple feature rankings would need to get serialized into a single feature rank list. reRank accomplishes this by calculating the rank sum for each feature across all "1 vs. all" feature selection tasks. Features found in one rank list but not in another receive a numeric rank equal to one more than the maximum rank in that feature rank list. The presence of a NA placeholder (see: doMulti) will not impact reRank.

We note here, however, that a better approach would deploy "1 vs. all" feature selection and classifier construction simultaneously, rather than "1 vs. all" feature selection followed by "1 vs. all" classifier construction. This is now implemented as plGridMulti.

A vector of re-ranked features. See Details.

exprso documentation built on May 1, 2019, 7:11 p.m.