R/extract_factors.R

In sqlscore: Utilities for Generating SQL Queries from Model Objects

# Standardized factor extraction
#
# Extract factor variables expanded by a model object's constructor
#
# Extract the names of factor variables passed to a model's constructor, and return
# them as a character vector. This is useful for linpred() in identifying expanded
# dummies so that they can be turned into CASE expressions.
#
# @param object An object for which the extraction of input factors is meaningful.
#
# @return Names of model input factors as a character vector.
extract_factors <-
function(object)
{
  UseMethod("extract_factors")
}

#' @export
extract_factors.default <-
function(object)
{
  nm <- attr(stats::terms(object), "dataClasses")
  factors <- names(nm[which(nm %in% c("factor", "character"))])

  # if no contrasts specification, assume all are contr.treatment
  if("contrasts" %in% names(object))
  {
    ct <- unique(sapply(factors, function(x) object$contrasts[[x]]))

    if(length(ct) > 0 && (length(ct) > 1 || ct != "contr.treatment"))
    {
      stop("Only treatment contrasts can be used for SQL generation")
    }
  }

  factors
}

#' @export
extract_factors.glmboost <-
function(object)
{
  cl <- as.list(object$call)
  if("formula" %in% names(cl))
  {
    dc <- attr(attr(object$model.frame(), 'terms', exact=TRUE),
               'dataClasses', exact=TRUE)
    av <- names(which(dc == "factor"))
    av <- av[2:length(av)]

    tp <- sapply(object$model.frame(), is.factor)
    tp <- names(tp[which(tp)])

    # it's poorly documented whether model.frame() includes
    # only vars used in the fit, so let's be safe
    return(intersect(tp, av))
  } else
  {
    # the fit was with x and y arrays, can't recover factors
    return(c())
  }
}

#' @export
extract_factors.cv.glmnet <-
function(object)
{
  # cv.glmnet's x argument has to be a pre-expanded matrix;
  # there's no way to recover any dummy expansion that was done,
  # so using this package with glmnet just requires that expansion
  # be done in the DB (and that the source table have columns for the
  # dummies)
  c()
}