Add ranger-based knockoff statistic

[ahrensmaike]

Scope

stat_random_forest() is currently based on the {randomForestSRC} package.
The {ranger} package is more commonly used than {randomForestSRC},
has arguably more sensible defaults (e.g. for importance measure),
can handle all three endpoint types and is fast
(C++ implementation).

• Implement new function(s) for {ranger}-based modelling
• Importance measure: permutation importance provided by {ranger}
• Write tests in analogy to stat_random_forest()

Details

The following description of individual components follows the current
structure in the package.

• stat_ranger <- function(X, X_k, y, type = "regression", ...) { ... }
  • input as for existing stat_* functions, with
    ... to pass further arguments for importance score calculation
    (can apply to modelling or importance, resp.)
  • output: data frame/tibble, exactly one column named W,
    one row per predictor, rownames equal to colnames(X)
• ranger_importance_scores() called within stat_ranger() to
  fit random forest and compute/extract individual importances for
  knockoff and original variables

Known issues

• in rare cases, ranger may return NaN for variable importance.
  We have seen this in our initial custom implementation of stat_ranger()
  and need to come up with an efficient solution.

Questions

• random forest options:
  • keep stat_random_forest() with {randomforestSRC} and add stat_ranger()?
  • deprecate {randomforestSRC}?
  • keep stat_random_forest() and add engine argument to select between "randomforestSRC" and "ranger"?

Related tasks

• update knockoff.statistics() documentation
• parallelization framework

[kormama1]

@ahrensmaike @svoss09 Regarding stat_random_forest, I would say let's not deprecate it (unless there was a critical bug). I would just downplay its importance.

E.g. here in the documentation of knockoff.statistics: "#' @param statistic knockoff feature statistic function, defaults to glmnet coefficient difference (statistic="stat_glmnet"; see ?stat_glmnet). Other options include statistic="stat_random_forest" (see ?stat_random_forest), statistic="stat_predictive_glmnet" (see ?stat_predictive_glmnet) or statistic="stat_predictive_causal_forest" (see ?stat_predictive_causal_forest)."

Instead we could start with: "Other options include statistic = "stat_ranger" (see ?stat_ranger) etc. ..." then we could end the listing of options with statistic="stat_random_forest" (which is a prior implementation of random forest importance score, if users prefer; in general, however, we recommend using statistic = "stat_ranger" for random forest implementation).

[kormama1]

Regarding engine argument, perhaps it's overly complex for such a minor thing and we could just (try to) make sure that users will always choose stat_ranger over the old implementation (unless they really wanted to use the older version) as per my prior comment above.

[kormama1]

Perhaps we can deprecate stat_random_forest at a later time point when we've had a few months of flawless usage experience with stat_ranger, if we feel it's simpler (which in principle it is).