# S3 method for gafs varImp(object, metric = object$control$metric["external"], maximize = object$control$maximize["external"], ...)
| object | |
|---|---|
| metric | a metric to compute importance (see Details below) |
| maximize | are larger values of the metric better? |
| ... | not currently uses |
a data frame where the rownames are the predictor names and the column is the average t-statistic
A crude measure of importance is computed for thee two search procedures. At
the end of a search process, the difference in the fitness values is
computed for models with and without each feature (based on the search
history). If a predictor has at least two subsets that include and did not
include the predictor, a t-statistic is computed (otherwise a value of
NA is assigned to the predictor).
This computation is done separately for each resample and the t-statistics
are averaged (NA values are ignored) and this average is reported as
the importance. If the fitness value should be minimized, the negative value
of the t-statistic is used in the average.
As such, the importance score reflects the standardized increase in fitness that occurs when the predict is included in the subset. Values near zero (or negative) indicate that the predictor may not be important to the model.