Nothing
R/compboost.R
In compboost: C++ Implementation of Component-Wise Boosting
#' Compboost API
#'
#' \code{Compboost} wraps the \code{S4} class system exposed by \code{Rcpp} to make defining
#' objects, adding objects, the training and taking predictions, and plotting much easier.
#' As already mentioned, the \code{Compboost} class is just a wrapper and hence compatible
#' with the most \code{S4} classes. This together defines the compboost API.
#'
#' @format \code{\link{R6Class}} object.
#' @name Compboost
#' @section Usage:
#' \preformatted{
#' cboost = Compboost$new(data, target, optimizer = OptimizerCoordinateDescent$new(), loss,
#' learning.rate = 0.05)
#'
#' cboost$addLogger(logger, use.as.stopper = FALSE, logger.id, ...)
#'
#' cbboost$addBaselearner(features, id, bl.factory, data.source = InMemoryData,
#' data.target = InMemoryData, ...)
#'
#' cbboost$train(iteration = 100, trace = TRUE)
#'
#' cboost$getCurrentIteration()
#'
#' cboost$predict(newdata = NULL)
#'
#' cboost$getInbagRisk()
#'
#' cboost$getSelectedBaselearner()
#'
#' cboost$getEstimatedCoef()
#'
#' cboost$plot(blearner.type = NULL, iters = NULL, from = NULL, to = NULL, length.out = 1000)
#'
#' cboost$getBaselearnerNames()
#'
#' cboost$prepareData(newdata)
#'
#' }
#' @section Arguments:
#' \strong{For Compboost$new()}:
#' \describe{
#' \item{\code{data}}{[\code{data.frame}]\cr
#' Data used for training.
#' }
#' \item{\code{target}}{[\code{character(1)}]\cr
#' Character naming the target. It is necessary that target is available as column in data.
#' }
#' \item{\code{optimizer}}{[\code{S4 Optimizer}]\cr
#' Optimizer used for the fitting process given as initialized \code{S4 Optimizer} class.
#' Default is the \code{OptimizerCoordinateDescent}.
#' }
#' \item{\code{loss}}{[\code{S4 Loss}]\cr
#' Loss as initialized \code{S4 Loss} which is used to calculate pseudo residuals and the
#' empirical risk. Note that the loss needs match the data type of the target variable.
#' See the details for possible choices.
#' }
#' \item{\code{learning.rage}}{[\code{numeric(1)}]\cr
#' Learning rate used to shrink estimated parameter in each iteration. The learning rate
#' remains constant during the training and has to be between 0 and 1.
#' }
#' }
#'
#' \strong{For cboost$addLogger()}:
#' \describe{
#' \item{\code{logger}}{[\code{S4 Logger}]\cr
#' Logger which are registered within a logger list. The objects must be given as uninitialized
#' \code{S4 Logger} class. See the details for possible choices.
#' }
#' \item{\code{use.as.stopper}}{[\code{logical(1)}]\cr
#' Logical indicating whether the new logger should also be used as stopper. Default value is
#' \code{FALSE}.
#' }
#' \item{\code{logger.id}}{[\code{character(1)}]\cr
#' Id of the new logger. This is necessary to e.g. register multiple risk logger.
#' }
#' \item{}{\code{...}\cr
#' Further arguments passed to the constructor of the \code{S4 Logger} class specified in
#' \code{logger}. For possible arguments see details or the help pages (e.g. \code{?LoggerIteration})
#' of the \code{S4} classes.
#' }
#' }
#'
#' \strong{For cboost$addBaselearner()}:
#' \describe{
#' \item{\code{features}}{[\code{character()}]\cr
#' Vector of column names which are used as input data matrix for a single base-learner. Note that not
#' every base-learner supports the use of multiple features (e.g. the spline base-learner).
#' }
#' \item{\code{id}}{[\code{character(1)}]\cr
#' Id of the base-learners. This is necessary since it is possible to define multiple learners with the same underlying data.
#' }
#' \item{\code{bl.factory}}{[\code{S4 Factory}]\cr
#' Uninitialized base-learner factory represented as \code{S4 Factory} class. See the details
#' for possible choices.
#' }
#' \item{\code{data.source}}{[\code{S4 Data}]\cr
#' Data source object. At the moment just in memory is supported.
#' }
#' \item{\code{data.target}}{[\code{S4 Data}]\cr
#' Data target object. At the moment just in memory is supported.
#' }
#' \item{}{\code{...}\cr
#' Further arguments passed to the constructor of the \code{S4 Factory} class specified in
#' \code{bl.factory}. For possible arguments see the help pages (e.g. \code{?BaselearnerPSplineFactory})
#' of the \code{S4} classes.
#' }
#' }
#'
#' \strong{For cboost$train()}:
#' \describe{
#' \item{\code{iteration}}{[\code{integer(1)}]\cr
#' Set the algorithm at \code{iteration}. Note: This argument is ignored if this is the first
#' training and an iteration logger is already specified. For further uses the algorithm automatically
#' continues training if \code{iteration} is set to an value larger than the already trained iterations.
#' }
#' \item{\code{trace}}{[\code{integer(1)}]\cr
#' Integer indicating how often a trace should be printed. Specifying \code{trace = 10}, then every
#' 10th iteration is printed. If no trace should be printed set \code{trace = 0}. Default is
#' -1 which means that we set \code{trace} at a value that 40 iterations are printed.
#' }
#' }
#'
#' \strong{For cboost$predict()}:
#' \describe{
#' \item{\code{newdata}}{[\code{data.frame()}]\cr
#' Data to predict on. If \code{NULL} predictions on the training data are returned.
#' }
#' }
#' \strong{For cboost$plot()}:
#' \describe{
#' \item{\code{blearner.type}}{[\code{character(1)}]\cr
#' Character name of the base-learner to plot the additional contribution to the response.
#' }
#' \item{\code{iters}}{[\code{integer()}]\cr
#' Integer vector containing the iterations the user wants to illustrate.
#' }
#' \item{\code{from}}{[\code{numeric(1)}]\cr
#' Lower bound for plotting (should be smaller than \code{to}).
#' }
#' \item{\code{to}}{[\code{numeric(1)}]\cr
#' Upper bound for plotting (should be greater than \code{from}).
#' }
#' \item{\code{length.out}}{[\code{integer(1)}]\cr
#' Number of equidistant points between \code{from} and \code{to} used for plotting.
#' }
#' }
#' @section Details:
#' \strong{Loss}\cr
#' Available choices for the loss are:
#' \itemize{
#' \item
#' \code{LossQuadratic} (Regression)
#'
#' \item
#' \code{LossAbsolute} (Regression)
#'
#' \item
#' \code{LossBinomial} (Binary Classification)
#'
#' \item
#' \code{LossCustom} (Custom)
#'
#' \item
#' \code{LossCustomCpp} (Custom)
#' }
#' (For each loss also take a look at the help pages (e.g. \code{?LossBinomial}) and the
#' \code{C++} documentation for details about the underlying formulas)
#'
#' \strong{Logger}\cr
#' Available choices for the logger are:
#' \itemize{
#' \item
#' \code{LoggerIteration}: Log current iteration. Additional arguments:
#' \describe{
#' \item{\code{max_iterations} [\code{integer(1)}]}{
#' Maximal number of iterations.
#' }
#' }
#'
#' \item
#' \code{LoggerTime}: Log already elapsed time. Additional arguments:
#' \describe{
#' \item{\code{max_time} [\code{integer(1)}]}{
#' Maximal time for the computation.
#' }
#' \item{\code{time_unit} [\code{character(1)}]}{
#' Character to specify the time unit. Possible choices are \code{minutes}, \code{seconds}, or \code{microseconds}.
#' }
#' }
#'
#' \item
#' \code{LoggerInbagRisk}:
#' \describe{
#' \item{\code{used_loss} [\code{S4 Loss}]}{
#' Loss as initialized \code{S4 Loss} which is used to calculate the empirical risk. See the
#' details for possible choices.
#' }
#' \item{\code{eps_for_break} [\code{numeric(1)}]}{
#' This argument is used if the logger is also used as stopper. If the relative improvement
#' of the logged inbag risk falls above this boundary the stopper breaks the algorithm.
#' }
#' }
#'
#' \item
#' \code{LoggerOobRisk}:
#' \describe{
#' \item{\code{used_loss} [\code{S4 Loss}]}{
#' Loss as initialized \code{S4 Loss} which is used to calculate the empirical risk. See the
#' details for possible choices.
#' }
#' \item{\code{eps_for_break} [\code{numeric(1)}]}{
#' This argument is used if the logger is also used as stopper. If the relative improvement
#' of the logged inbag risk falls above this boundary the stopper breaks the algorithm.
#' }
#' \item{\code{oob_data} [\code{list}]}{
#' A list which contains data source objects which corresponds to the source data of each registered factory.
#' The source data objects should contain the out of bag data. This data is then used to calculate the
#' new predictions in each iteration.
#' }
#' \item{\code{oob_response} [\code{vector}]}{
#' Vector which contains the response for the out of bag data given within \code{oob_data}.
#' }
#' }
#' }
#' \strong{Note}:
#' \itemize{
#' \item
#' Even if you do not use the logger as stopper you have to define the arguments such as \code{max_time}.
#'
#' \item
#' We are aware of that the style guide here is not consistent with the \code{R6} arguments. Nevertheless, using
#' \code{_} as word separator is due to the used arguments within \code{C++}.
#' }
#'
#' @section Fields:
#' \describe{
#' \item{\code{data} [\code{data.frame}]}{
#' Data used for training the algorithm.
#' }
#' \item{\code{response} [\code{vector}]}{
#' Response given as vector.
#' }
#' \item{\code{target} [\code{character(1)}]}{
#' Name of the Response.
#' }
#' \item{\code{id} [\code{character(1)}]}{
#' Value to identify the data. By default name of \code{data}, but can be overwritten.
#' }
#' \item{\code{optimizer} [\code{S4 Optimizer}]}{
#' Optimizer used within the fitting process.
#' }
#' \item{\code{loss} [\code{S4 Loss}]}{
#' Loss used to calculate pseudo residuals and empirical risk.
#' }
#' \item{\code{learning.rate} [\code{numeric(1)}]}{
#' Learning rate used to shrink the estimated parameter in each iteration.
#' }
#' \item{\code{model} [\code{S4 Compboost_internal}]}{
#' Internal \code{S4 Compboost_internal} class on which the main operations are called.
#' }
#' \item{\code{bl.factory.list} [\code{S4 FactoryList}]}{
#' List of all registered factories represented as \code{S4 FactoryList} class.
#' }
#' \item{\code{positive.category} [\code{character(1)}]}{
#' Character containing the name of the positive class in the case of classification.
#' }
#' \item{\code{stop.if.all.stoppers.fulfilled} [\code{logical(1)}]}{
#' Logical indicating whether all stopper should be used simultaneously or if it is sufficient
#' that the first stopper which is fulfilled breaks the algorithm.
#' }
#' }
#'
#' @section Methods:
#' \describe{
#' \item{\code{addLogger}}{method to add a logger to the algorithm (Note: This is just possible before the training).}
#' \item{\code{addBaselearner}}{method to add a new base-learner factories to the algorithm (Note: This is just possible before the training).}
#' \item{\code{getCurrentIteration}}{method to get the current iteration on which the algorithm is set.}
#' \item{\code{train}}{method to train the algorithm.}
#' \item{\code{predict}}{method to predict on a trained object.}
#' \item{\code{getSelectedBaselearner}}{method to get a character vector of selected base-learner.}
#' \item{\code{getEstimatedCoef}}{method to get a list of estimated coefficient for each selected base-learner.}
#' \item{\code{plot}}{method to plot the \code{Compboost} object.}
#' \item{\code{getBaselearnerNames}}{method to get names of registered factories.}
#' }
#'
#' @examples
#' cboost = Compboost$new(mtcars, "mpg", loss = LossQuadratic$new())
#' cboost$addBaselearner("hp", "spline", BaselearnerPSpline, degree = 3,
#' n.knots = 10, penalty = 2, differences = 2)
#' cboost$train(1000)
#'
#' table(cboost$getSelectedBaselearner())
#' cboost$plot("hp_spline")
NULL
#'@export
Compboost = R6::R6Class("Compboost",
public = list(
data = NULL,
response = NULL,
target = NULL,
id = NULL,
optimizer = NULL,
loss = NULL,
learning.rate = NULL,
model = NULL,
bl.factory.list = NULL,
positive.category = NULL,
stop.if.all.stoppers.fulfilled = FALSE,
initialize = function(data, target, optimizer = OptimizerCoordinateDescent$new(), loss, learning.rate = 0.05) {
checkmate::assertDataFrame(data, any.missing = FALSE, min.rows = 1)
checkmate::assertCharacter(target)
checkmate::assertNumeric(learning.rate, lower = 0, upper = 1, len = 1)
if (! target %in% names(data)) {
stop ("The target ", target, " is not present within the data")
}
if (inherits(loss, "C++Class")) {
stop ("Loss should be an initialized loss object by calling the constructor: ", deparse(substitute(loss)), "$new()")
}
self$id = deparse(substitute(data))
data = droplevels(as.data.frame(data))
response = data[[target]]
# Transform factor or character labels to -1 and 1
if (! is.numeric(response)) {
response = as.factor(response)
if (length(levels(response)) > 2) {
stop("Multiclass classification is not supported.")
}
self$positive.category = levels(response)[1]
# Transform to vector with -1 and 1:
response = as.integer(response) * (1 - as.integer(response)) + 1
}
self$target = target
self$response = response
self$data = data[, !colnames(data) %in% target, drop = FALSE]
self$optimizer = optimizer
self$loss = loss
self$learning.rate = learning.rate
# Initialize new base-learner factory list. All factories which are defined in
# `addBaselearners` are registered here:
self$bl.factory.list = BlearnerFactoryList$new()
},
addLogger = function(logger, use.as.stopper = FALSE, logger.id, ...) {
private$l.list[[logger.id]] = logger$new(use.as.stopper = use.as.stopper, ...)
},
getCurrentIteration = function() {
if (!is.null(self$model) && self$model$isTrained()) {
return(length(self$model$getSelectedBaselearner()))
} else {
return(0)
}
},
addBaselearner = function(feature, id, bl.factory, data.source = InMemoryData, data.target = InMemoryData, ...) {
if (!is.null(self$model)) {
stop("No base-learners can be added after training is started")
}
# Clear base-learners which are within the bl.list but not registered:
idx.remove = ! names(private$bl.list) %in% self$bl.factory.list$getRegisteredFactoryNames()
if (any(idx.remove)) {
for (i in which(idx.remove)) {
private$bl.list[[i]] = NULL
}
}
data.columns = self$data[, feature, drop = FALSE]
id.fac = paste(paste(feature, collapse = "_"), id, sep = "_") #USE stringi
if (ncol(data.columns) == 1 && !is.numeric(data.columns[, 1])) {
private$addSingleCatBl(data.columns, feature, id, id.fac, bl.factory, data.source, data.target, ...)
} else {
private$addSingleNumericBl(data.columns, feature, id, id.fac, bl.factory, data.source, data.target, ...)
}
},
train = function(iteration = 100, trace = -1) {
if (self$bl.factory.list$getNumberOfRegisteredFactories() == 0) {
stop("Could not train without any registered base-learner.")
}
checkmate::assertIntegerish(iteration, lower = 1, len = 1, null.ok = TRUE)
# checkmate::assertFlag(trace)
checkmate::assertIntegerish(trace, lower = -1, upper = iteration, len = 1, null.ok = FALSE)
if (trace == -1) {
trace = round(iteration / 40)
}
# Check if it is necessary to add a initial iteration logger. This is not the case
# when the user already has add one by calling `addLogger`:
if (is.null(self$model)) {
# If iteration is NULL, then there is no new iteration logger defined. This could be
# used, for example, to train the algorithm an break it after a defined number of
# hours or minutes.
if (!is.null(iteration)) {
# Add new logger in the case that there isn't already a custom defined one:
if ("Rcpp_LoggerIteration" %in% vapply(private$l.list, class, character(1))) {
warning("Training iterations are ignored since custom iteration logger is already defined")
} else {
self$addLogger(LoggerIteration, TRUE, logger.id = "_iterations", iter.max = iteration)
}
}
# After calling `initializeModel` it isn't possible to add base-learner or logger.
private$initializeModel()
}
# Just call train for the initial fitting process. If the model is alredy initially trained,
# then we use `setToIteration` to set to a lower iteration or retrain the model.
if (!self$model$isTrained())
self$model$train(trace)
else {
self$model$setToIteration(iteration)
}
return(invisible(NULL))
},
prepareData = function (newdata) {
new.source.features = unique(lapply(private$bl.list, function (x) x$feature))
new.sources = list()
data.names = character()
# Remove lapply due to categorical feature handling which needs to return multiple data objects
# at once.
for (ns in new.source.features) {
data.columns = newdata[, ns, drop = FALSE]
if (ncol(data.columns) == 1 && !is.numeric(data.columns[, 1])) {
lvls = unlist(unique(data.columns))
# Create dummy variable for each category and use that vector as data matrix. Hence,
# if a categorical feature has 3 groups, then these 3 groups are added as 3 different
# base-learners (unbiased feature selection).
for (lvl in lvls) {
data.names = append(data.names, paste(ns, lvl, sep = "_"))
new.sources = c(new.sources, InMemoryData$new(as.matrix(as.integer(data.columns == lvl)), paste(ns, lvl, sep = "_")))
}
} else {
data.names = append(data.names, paste(ns, collapse = "_"))
new.sources = c(new.sources, InMemoryData$new(as.matrix(data.columns), paste(ns, collapse = "_")))
}
}
names(new.sources) = data.names
return(new.sources)
},
predict = function(newdata = NULL, response = FALSE) {
checkmate::assertDataFrame(newdata, null.ok = TRUE, min.rows = 1)
if (is.null(newdata)) {
return(self$model$getPrediction(response))
} else {
return(self$model$predict(self$prepareData(newdata), response))
}
},
getInbagRisk = function() {
if(!is.null(self$model)) {
# Return the risk + intercept, hence the current iteration + 1:
return(self$model$getRiskVector()[seq_len(self$getCurrentIteration() + 1)])
}
return(NULL)
},
getSelectedBaselearner = function() {
if(!is.null(self$model))
return(self$model$getSelectedBaselearner())
return(NULL)
},
print = function() {
p = glue::glue("\n
Component-Wise Gradient Boosting\n
Trained on {self$id} with target {self$target}
Number of base-learners: {self$bl.factory.list$getNumberOfRegisteredFactories()}
Learning rate: {self$learning.rate}
Iterations: {self$getCurrentIteration()}
")
if (! is.null(self$positive.category))
p = glue::glue(p, "\nPositive class: {self$positive.category}")
if(!is.null(self$model))
p = glue::glue(p, "\nOffset: {round(self$model$getOffset(), 4)}")
print(p)
print(self$loss)
},
getEstimatedCoef = function () {
if(!is.null(self$model)) {
return(c(self$model$getEstimatedParameter(), offset = self$model$getOffset()))
}
return(NULL)
},
plot = function (blearner.type = NULL, iters = NULL, from = NULL, to = NULL, length.out = 1000) {
if (requireNamespace("ggplot2", quietly = TRUE)) {
if (is.null(self$model)) {
stop("Model needs to be trained first.")
}
checkmate::assertIntegerish(iters, min.len = 1, any.missing = FALSE, null.ok = TRUE)
checkmate::assertCharacter(blearner.type, len = 1, null.ok = TRUE)
if (is.null(blearner.type)) {
stop("Please specify a valid base-learner plus feature.")
}
if (! blearner.type %in% names(private$bl.list)) {
stop("Your requested feature plus learner is not available. Check 'getBaselearnerNames()' for available learners.")
}
if (length(private$bl.list[[blearner.type]]$feature) > 1) {
stop("Only univariate plotting is supported.")
}
# Check if selected base-learner includes the proposed one + check if iters is big enough:
iter.min = which(self$getSelectedBaselearner() == blearner.type)[1]
if (! blearner.type %in% unique(self$getSelectedBaselearner())) {
stop("Requested base-learner plus feature was not selected.")
} else {
if (any(iters < iter.min)) {
warning("Requested base-learner plus feature was first selected at iteration ", iter.min)
}
}
feat.name = private$bl.list[[blearner.type]]$target$getIdentifier()
checkmate::assertNumeric(x = self$data[[feat.name]], min.len = 2, null.ok = FALSE)
checkmate::assertNumeric(from, lower = min(self$data[[feat.name]]), upper = max(self$data[[feat.name]]), len = 1, null.ok = TRUE)
checkmate::assertNumeric(to, lower = min(self$data[[feat.name]]), upper = max(self$data[[feat.name]]), len = 1, null.ok = TRUE)
if (is.null(from)) {
from = min(self$data[[feat.name]])
}
if (is.null(to)) {
to = max(self$data[[feat.name]])
}
if (from >= to) {
warning("Argument from is smaller than to, hence the x interval is [to, from].")
temp = from
from = to
to = temp
}
plot.data = as.matrix(seq(from = from, to = to, length.out = length.out))
feat.map = private$bl.list[[blearner.type]]$factory$transformData(plot.data)
# Create data.frame for plotting depending if iters is specified:
if (!is.null(iters[1])) {
preds = lapply(iters, function (x) {
if (x >= iter.min) {
return(feat.map %*% self$model$getParameterAtIteration(x)[[blearner.type]])
} else {
return(rep(0, length.out))
}
})
names(preds) = iters
df.plot = data.frame(
effect = unlist(preds),
iteration = as.factor(rep(iters, each = length.out)),
feature = plot.data
)
gg = ggplot2::ggplot(df.plot, ggplot2::aes(feature, effect, color = iteration))
} else {
df.plot = data.frame(
effect = feat.map %*% self$getEstimatedCoef()[[blearner.type]],
feature = plot.data
)
gg = ggplot2::ggplot(df.plot, ggplot2::aes(feature, effect))
}
# If there are too much rows we need to take just a sample or completely remove rugs:
if (nrow(self$data) > 1000) {
idx.rugs = sample(seq_len(nrow(self$data)), 1000, FALSE)
} else {
idx.rugs = seq_len(nrow(self$data))
}
gg = gg +
ggplot2::geom_line() +
ggplot2::geom_rug(data = self$data[idx.rugs,], ggplot2::aes_string(x = feat.name), inherit.aes = FALSE,
alpha = 0.8) +
ggplot2::xlab(feat.name) +
ggplot2::xlim(from, to) +
ggplot2::ylab("Additive Contribution") +
ggplot2::labs(title = paste0("Effect of ", blearner.type),
subtitle = "Additive contribution of predictor")
return(gg)
} else {
message("Please install ggplot2 to create plots.")
return(NULL)
}
},
getBaselearnerNames = function () {
# return(lapply(private$bl.list, function (bl) bl[[1]]$target$getIdentifier()))
return(names(private$bl.list))
}
),
private = list(
# Lists of single logger and base-learner factories. Necessary to prevent the factories from the
# arbage collector which deallocates all the data from the heap and couses R to crash.
l.list = list(),
bl.list = list(),
logger.list = list(),
initializeModel = function() {
private$logger.list = LoggerList$new()
for (n in names(private$l.list)) {
private$logger.list$registerLogger(n, private$l.list[[n]])
}
self$model = Compboost_internal$new(self$response, self$learning.rate,
self$stop.if.all.stoppers.fulfilled, self$bl.factory.list, self$loss, private$logger.list, self$optimizer)
},
addSingleNumericBl = function(data.columns, feature, id.fac, id, bl.factory, data.source, data.target, ...) {
private$bl.list[[id]] = list()
private$bl.list[[id]]$source = data.source$new(as.matrix(data.columns), paste(feature, collapse = "_"))
private$bl.list[[id]]$feature = feature
private$bl.list[[id]]$target = data.target$new()
# Call handler for default arguments and argument handling:
handler.name = paste0(".handle", bl.factory@.Data)
par.set = c(source = private$bl.list[[id]]$source, target = private$bl.list[[id]]$target, id = id.fac, do.call(handler.name, list(...)))
private$bl.list[[id]]$factory = do.call(bl.factory$new, par.set)
# private$bl.list[[id]]$factory = bl.factory$new(private$bl.list[[id]]$source, private$bl.list[[id]]$target, id.fac, ...)
self$bl.factory.list$registerFactory(private$bl.list[[id]]$factory)
private$bl.list[[id]]$source = NULL
},
addSingleCatBl = function(data.column, feature, id.fac, id, bl.factory, data.source, data.target, ...) {
lvls = unlist(unique(data.column))
# Create dummy variable for each category and use that vector as data matrix. Hence,
# if a categorical feature has 3 groups, then these 3 groups are added as 3 different
# base-learners (unbiased feature selection).
for (lvl in lvls) {
list.id = paste(feature, lvl, id.fac, sep = "_")
private$addSingleNumericBl(data.columns = as.matrix(as.integer(data.column == lvl)),
feature = paste(feature, lvl, sep = "_"), id.fac = id.fac,
id = list.id, bl.factory, data.source, data.target, ...)
# This is important because of:
# 1. feature in addSingleNumericBl needs to be something like cat_feature_Group1 to define the
# data objects correctly in a unique way.
# 2. The feature itself should not be named with the level. Instead of that we just want the
# feature name of the categorical variable, such as cat_feature (important for predictions).
private$bl.list[[list.id]]$feature = feature
}
}
)
)
Try the compboost package in your browser
Any scripts or data that you put into this service are public.
compboost documentation built on May 2, 2019, 6:40 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
#' Compboost API
#'
#' \code{Compboost} wraps the \code{S4} class system exposed by \code{Rcpp} to make defining
#' objects, adding objects, the training and taking predictions, and plotting much easier.
#' As already mentioned, the \code{Compboost} class is just a wrapper and hence compatible
#' with the most \code{S4} classes. This together defines the compboost API.
#'
#' @format \code{\link{R6Class}} object.
#' @name Compboost
#' @section Usage:
#' \preformatted{
#' cboost = Compboost$new(data, target, optimizer = OptimizerCoordinateDescent$new(), loss,
#' learning.rate = 0.05)
#'
#' cboost$addLogger(logger, use.as.stopper = FALSE, logger.id, ...)
#'
#' cbboost$addBaselearner(features, id, bl.factory, data.source = InMemoryData,
#' data.target = InMemoryData, ...)
#'
#' cbboost$train(iteration = 100, trace = TRUE)
#'
#' cboost$getCurrentIteration()
#'
#' cboost$predict(newdata = NULL)
#'
#' cboost$getInbagRisk()
#'
#' cboost$getSelectedBaselearner()
#'
#' cboost$getEstimatedCoef()
#'
#' cboost$plot(blearner.type = NULL, iters = NULL, from = NULL, to = NULL, length.out = 1000)
#'
#' cboost$getBaselearnerNames()
#'
#' cboost$prepareData(newdata)
#'
#' }
#' @section Arguments:
#' \strong{For Compboost$new()}:
#' \describe{
#' \item{\code{data}}{[\code{data.frame}]\cr
#' Data used for training.
#' }
#' \item{\code{target}}{[\code{character(1)}]\cr
#' Character naming the target. It is necessary that target is available as column in data.
#' }
#' \item{\code{optimizer}}{[\code{S4 Optimizer}]\cr
#' Optimizer used for the fitting process given as initialized \code{S4 Optimizer} class.
#' Default is the \code{OptimizerCoordinateDescent}.
#' }
#' \item{\code{loss}}{[\code{S4 Loss}]\cr
#' Loss as initialized \code{S4 Loss} which is used to calculate pseudo residuals and the
#' empirical risk. Note that the loss needs match the data type of the target variable.
#' See the details for possible choices.
#' }
#' \item{\code{learning.rage}}{[\code{numeric(1)}]\cr
#' Learning rate used to shrink estimated parameter in each iteration. The learning rate
#' remains constant during the training and has to be between 0 and 1.
#' }
#' }
#'
#' \strong{For cboost$addLogger()}:
#' \describe{
#' \item{\code{logger}}{[\code{S4 Logger}]\cr
#' Logger which are registered within a logger list. The objects must be given as uninitialized
#' \code{S4 Logger} class. See the details for possible choices.
#' }
#' \item{\code{use.as.stopper}}{[\code{logical(1)}]\cr
#' Logical indicating whether the new logger should also be used as stopper. Default value is
#' \code{FALSE}.
#' }
#' \item{\code{logger.id}}{[\code{character(1)}]\cr
#' Id of the new logger. This is necessary to e.g. register multiple risk logger.
#' }
#' \item{}{\code{...}\cr
#' Further arguments passed to the constructor of the \code{S4 Logger} class specified in
#' \code{logger}. For possible arguments see details or the help pages (e.g. \code{?LoggerIteration})
#' of the \code{S4} classes.
#' }
#' }
#'
#' \strong{For cboost$addBaselearner()}:
#' \describe{
#' \item{\code{features}}{[\code{character()}]\cr
#' Vector of column names which are used as input data matrix for a single base-learner. Note that not
#' every base-learner supports the use of multiple features (e.g. the spline base-learner).
#' }
#' \item{\code{id}}{[\code{character(1)}]\cr
#' Id of the base-learners. This is necessary since it is possible to define multiple learners with the same underlying data.
#' }
#' \item{\code{bl.factory}}{[\code{S4 Factory}]\cr
#' Uninitialized base-learner factory represented as \code{S4 Factory} class. See the details
#' for possible choices.
#' }
#' \item{\code{data.source}}{[\code{S4 Data}]\cr
#' Data source object. At the moment just in memory is supported.
#' }
#' \item{\code{data.target}}{[\code{S4 Data}]\cr
#' Data target object. At the moment just in memory is supported.
#' }
#' \item{}{\code{...}\cr
#' Further arguments passed to the constructor of the \code{S4 Factory} class specified in
#' \code{bl.factory}. For possible arguments see the help pages (e.g. \code{?BaselearnerPSplineFactory})
#' of the \code{S4} classes.
#' }
#' }
#'
#' \strong{For cboost$train()}:
#' \describe{
#' \item{\code{iteration}}{[\code{integer(1)}]\cr
#' Set the algorithm at \code{iteration}. Note: This argument is ignored if this is the first
#' training and an iteration logger is already specified. For further uses the algorithm automatically
#' continues training if \code{iteration} is set to an value larger than the already trained iterations.
#' }
#' \item{\code{trace}}{[\code{integer(1)}]\cr
#' Integer indicating how often a trace should be printed. Specifying \code{trace = 10}, then every
#' 10th iteration is printed. If no trace should be printed set \code{trace = 0}. Default is
#' -1 which means that we set \code{trace} at a value that 40 iterations are printed.
#' }
#' }
#'
#' \strong{For cboost$predict()}:
#' \describe{
#' \item{\code{newdata}}{[\code{data.frame()}]\cr
#' Data to predict on. If \code{NULL} predictions on the training data are returned.
#' }
#' }
#' \strong{For cboost$plot()}:
#' \describe{
#' \item{\code{blearner.type}}{[\code{character(1)}]\cr
#' Character name of the base-learner to plot the additional contribution to the response.
#' }
#' \item{\code{iters}}{[\code{integer()}]\cr
#' Integer vector containing the iterations the user wants to illustrate.
#' }
#' \item{\code{from}}{[\code{numeric(1)}]\cr
#' Lower bound for plotting (should be smaller than \code{to}).
#' }
#' \item{\code{to}}{[\code{numeric(1)}]\cr
#' Upper bound for plotting (should be greater than \code{from}).
#' }
#' \item{\code{length.out}}{[\code{integer(1)}]\cr
#' Number of equidistant points between \code{from} and \code{to} used for plotting.
#' }
#' }
#' @section Details:
#' \strong{Loss}\cr
#' Available choices for the loss are:
#' \itemize{
#' \item
#' \code{LossQuadratic} (Regression)
#'
#' \item
#' \code{LossAbsolute} (Regression)
#'
#' \item
#' \code{LossBinomial} (Binary Classification)
#'
#' \item
#' \code{LossCustom} (Custom)
#'
#' \item
#' \code{LossCustomCpp} (Custom)
#' }
#' (For each loss also take a look at the help pages (e.g. \code{?LossBinomial}) and the
#' \code{C++} documentation for details about the underlying formulas)
#'
#' \strong{Logger}\cr
#' Available choices for the logger are:
#' \itemize{
#' \item
#' \code{LoggerIteration}: Log current iteration. Additional arguments:
#' \describe{
#' \item{\code{max_iterations} [\code{integer(1)}]}{
#' Maximal number of iterations.
#' }
#' }
#'
#' \item
#' \code{LoggerTime}: Log already elapsed time. Additional arguments:
#' \describe{
#' \item{\code{max_time} [\code{integer(1)}]}{
#' Maximal time for the computation.
#' }
#' \item{\code{time_unit} [\code{character(1)}]}{
#' Character to specify the time unit. Possible choices are \code{minutes}, \code{seconds}, or \code{microseconds}.
#' }
#' }
#'
#' \item
#' \code{LoggerInbagRisk}:
#' \describe{
#' \item{\code{used_loss} [\code{S4 Loss}]}{
#' Loss as initialized \code{S4 Loss} which is used to calculate the empirical risk. See the
#' details for possible choices.
#' }
#' \item{\code{eps_for_break} [\code{numeric(1)}]}{
#' This argument is used if the logger is also used as stopper. If the relative improvement
#' of the logged inbag risk falls above this boundary the stopper breaks the algorithm.
#' }
#' }
#'
#' \item
#' \code{LoggerOobRisk}:
#' \describe{
#' \item{\code{used_loss} [\code{S4 Loss}]}{
#' Loss as initialized \code{S4 Loss} which is used to calculate the empirical risk. See the
#' details for possible choices.
#' }
#' \item{\code{eps_for_break} [\code{numeric(1)}]}{
#' This argument is used if the logger is also used as stopper. If the relative improvement
#' of the logged inbag risk falls above this boundary the stopper breaks the algorithm.
#' }
#' \item{\code{oob_data} [\code{list}]}{
#' A list which contains data source objects which corresponds to the source data of each registered factory.
#' The source data objects should contain the out of bag data. This data is then used to calculate the
#' new predictions in each iteration.
#' }
#' \item{\code{oob_response} [\code{vector}]}{
#' Vector which contains the response for the out of bag data given within \code{oob_data}.
#' }
#' }
#' }
#' \strong{Note}:
#' \itemize{
#' \item
#' Even if you do not use the logger as stopper you have to define the arguments such as \code{max_time}.
#'
#' \item
#' We are aware of that the style guide here is not consistent with the \code{R6} arguments. Nevertheless, using
#' \code{_} as word separator is due to the used arguments within \code{C++}.
#' }
#'
#' @section Fields:
#' \describe{
#' \item{\code{data} [\code{data.frame}]}{
#' Data used for training the algorithm.
#' }
#' \item{\code{response} [\code{vector}]}{
#' Response given as vector.
#' }
#' \item{\code{target} [\code{character(1)}]}{
#' Name of the Response.
#' }
#' \item{\code{id} [\code{character(1)}]}{
#' Value to identify the data. By default name of \code{data}, but can be overwritten.
#' }
#' \item{\code{optimizer} [\code{S4 Optimizer}]}{
#' Optimizer used within the fitting process.
#' }
#' \item{\code{loss} [\code{S4 Loss}]}{
#' Loss used to calculate pseudo residuals and empirical risk.
#' }
#' \item{\code{learning.rate} [\code{numeric(1)}]}{
#' Learning rate used to shrink the estimated parameter in each iteration.
#' }
#' \item{\code{model} [\code{S4 Compboost_internal}]}{
#' Internal \code{S4 Compboost_internal} class on which the main operations are called.
#' }
#' \item{\code{bl.factory.list} [\code{S4 FactoryList}]}{
#' List of all registered factories represented as \code{S4 FactoryList} class.
#' }
#' \item{\code{positive.category} [\code{character(1)}]}{
#' Character containing the name of the positive class in the case of classification.
#' }
#' \item{\code{stop.if.all.stoppers.fulfilled} [\code{logical(1)}]}{
#' Logical indicating whether all stopper should be used simultaneously or if it is sufficient
#' that the first stopper which is fulfilled breaks the algorithm.
#' }
#' }
#'
#' @section Methods:
#' \describe{
#' \item{\code{addLogger}}{method to add a logger to the algorithm (Note: This is just possible before the training).}
#' \item{\code{addBaselearner}}{method to add a new base-learner factories to the algorithm (Note: This is just possible before the training).}
#' \item{\code{getCurrentIteration}}{method to get the current iteration on which the algorithm is set.}
#' \item{\code{train}}{method to train the algorithm.}
#' \item{\code{predict}}{method to predict on a trained object.}
#' \item{\code{getSelectedBaselearner}}{method to get a character vector of selected base-learner.}
#' \item{\code{getEstimatedCoef}}{method to get a list of estimated coefficient for each selected base-learner.}
#' \item{\code{plot}}{method to plot the \code{Compboost} object.}
#' \item{\code{getBaselearnerNames}}{method to get names of registered factories.}
#' }
#'
#' @examples
#' cboost = Compboost$new(mtcars, "mpg", loss = LossQuadratic$new())
#' cboost$addBaselearner("hp", "spline", BaselearnerPSpline, degree = 3,
#' n.knots = 10, penalty = 2, differences = 2)
#' cboost$train(1000)
#'
#' table(cboost$getSelectedBaselearner())
#' cboost$plot("hp_spline")
NULL
#'@export
Compboost = R6::R6Class("Compboost",
public = list(
data = NULL,
response = NULL,
target = NULL,
id = NULL,
optimizer = NULL,
loss = NULL,
learning.rate = NULL,
model = NULL,
bl.factory.list = NULL,
positive.category = NULL,
stop.if.all.stoppers.fulfilled = FALSE,
initialize = function(data, target, optimizer = OptimizerCoordinateDescent$new(), loss, learning.rate = 0.05) {
checkmate::assertDataFrame(data, any.missing = FALSE, min.rows = 1)
checkmate::assertCharacter(target)
checkmate::assertNumeric(learning.rate, lower = 0, upper = 1, len = 1)
if (! target %in% names(data)) {
stop ("The target ", target, " is not present within the data")
}
if (inherits(loss, "C++Class")) {
stop ("Loss should be an initialized loss object by calling the constructor: ", deparse(substitute(loss)), "$new()")
}
self$id = deparse(substitute(data))
data = droplevels(as.data.frame(data))
response = data[[target]]
# Transform factor or character labels to -1 and 1
if (! is.numeric(response)) {
response = as.factor(response)
if (length(levels(response)) > 2) {
stop("Multiclass classification is not supported.")
}
self$positive.category = levels(response)[1]
# Transform to vector with -1 and 1:
response = as.integer(response) * (1 - as.integer(response)) + 1
}
self$target = target
self$response = response
self$data = data[, !colnames(data) %in% target, drop = FALSE]
self$optimizer = optimizer
self$loss = loss
self$learning.rate = learning.rate
# Initialize new base-learner factory list. All factories which are defined in
# `addBaselearners` are registered here:
self$bl.factory.list = BlearnerFactoryList$new()
},
addLogger = function(logger, use.as.stopper = FALSE, logger.id, ...) {
private$l.list[[logger.id]] = logger$new(use.as.stopper = use.as.stopper, ...)
},
getCurrentIteration = function() {
if (!is.null(self$model) && self$model$isTrained()) {
return(length(self$model$getSelectedBaselearner()))
} else {
return(0)
}
},
addBaselearner = function(feature, id, bl.factory, data.source = InMemoryData, data.target = InMemoryData, ...) {
if (!is.null(self$model)) {
stop("No base-learners can be added after training is started")
}
# Clear base-learners which are within the bl.list but not registered:
idx.remove = ! names(private$bl.list) %in% self$bl.factory.list$getRegisteredFactoryNames()
if (any(idx.remove)) {
for (i in which(idx.remove)) {
private$bl.list[[i]] = NULL
}
}
data.columns = self$data[, feature, drop = FALSE]
id.fac = paste(paste(feature, collapse = "_"), id, sep = "_") #USE stringi
if (ncol(data.columns) == 1 && !is.numeric(data.columns[, 1])) {
private$addSingleCatBl(data.columns, feature, id, id.fac, bl.factory, data.source, data.target, ...)
} else {
private$addSingleNumericBl(data.columns, feature, id, id.fac, bl.factory, data.source, data.target, ...)
}
},
train = function(iteration = 100, trace = -1) {
if (self$bl.factory.list$getNumberOfRegisteredFactories() == 0) {
stop("Could not train without any registered base-learner.")
}
checkmate::assertIntegerish(iteration, lower = 1, len = 1, null.ok = TRUE)
# checkmate::assertFlag(trace)
checkmate::assertIntegerish(trace, lower = -1, upper = iteration, len = 1, null.ok = FALSE)
if (trace == -1) {
trace = round(iteration / 40)
}
# Check if it is necessary to add a initial iteration logger. This is not the case
# when the user already has add one by calling `addLogger`:
if (is.null(self$model)) {
# If iteration is NULL, then there is no new iteration logger defined. This could be
# used, for example, to train the algorithm an break it after a defined number of
# hours or minutes.
if (!is.null(iteration)) {
# Add new logger in the case that there isn't already a custom defined one:
if ("Rcpp_LoggerIteration" %in% vapply(private$l.list, class, character(1))) {
warning("Training iterations are ignored since custom iteration logger is already defined")
} else {
self$addLogger(LoggerIteration, TRUE, logger.id = "_iterations", iter.max = iteration)
}
}
# After calling `initializeModel` it isn't possible to add base-learner or logger.
private$initializeModel()
}
# Just call train for the initial fitting process. If the model is alredy initially trained,
# then we use `setToIteration` to set to a lower iteration or retrain the model.
if (!self$model$isTrained())
self$model$train(trace)
else {
self$model$setToIteration(iteration)
}
return(invisible(NULL))
},
prepareData = function (newdata) {
new.source.features = unique(lapply(private$bl.list, function (x) x$feature))
new.sources = list()
data.names = character()
# Remove lapply due to categorical feature handling which needs to return multiple data objects
# at once.
for (ns in new.source.features) {
data.columns = newdata[, ns, drop = FALSE]
if (ncol(data.columns) == 1 && !is.numeric(data.columns[, 1])) {
lvls = unlist(unique(data.columns))
# Create dummy variable for each category and use that vector as data matrix. Hence,
# if a categorical feature has 3 groups, then these 3 groups are added as 3 different
# base-learners (unbiased feature selection).
for (lvl in lvls) {
data.names = append(data.names, paste(ns, lvl, sep = "_"))
new.sources = c(new.sources, InMemoryData$new(as.matrix(as.integer(data.columns == lvl)), paste(ns, lvl, sep = "_")))
}
} else {
data.names = append(data.names, paste(ns, collapse = "_"))
new.sources = c(new.sources, InMemoryData$new(as.matrix(data.columns), paste(ns, collapse = "_")))
}
}
names(new.sources) = data.names
return(new.sources)
},
predict = function(newdata = NULL, response = FALSE) {
checkmate::assertDataFrame(newdata, null.ok = TRUE, min.rows = 1)
if (is.null(newdata)) {
return(self$model$getPrediction(response))
} else {
return(self$model$predict(self$prepareData(newdata), response))
}
},
getInbagRisk = function() {
if(!is.null(self$model)) {
# Return the risk + intercept, hence the current iteration + 1:
return(self$model$getRiskVector()[seq_len(self$getCurrentIteration() + 1)])
}
return(NULL)
},
getSelectedBaselearner = function() {
if(!is.null(self$model))
return(self$model$getSelectedBaselearner())
return(NULL)
},
print = function() {
p = glue::glue("\n
Component-Wise Gradient Boosting\n
Trained on {self$id} with target {self$target}
Number of base-learners: {self$bl.factory.list$getNumberOfRegisteredFactories()}
Learning rate: {self$learning.rate}
Iterations: {self$getCurrentIteration()}
")
if (! is.null(self$positive.category))
p = glue::glue(p, "\nPositive class: {self$positive.category}")
if(!is.null(self$model))
p = glue::glue(p, "\nOffset: {round(self$model$getOffset(), 4)}")
print(p)
print(self$loss)
},
getEstimatedCoef = function () {
if(!is.null(self$model)) {
return(c(self$model$getEstimatedParameter(), offset = self$model$getOffset()))
}
return(NULL)
},
plot = function (blearner.type = NULL, iters = NULL, from = NULL, to = NULL, length.out = 1000) {
if (requireNamespace("ggplot2", quietly = TRUE)) {
if (is.null(self$model)) {
stop("Model needs to be trained first.")
}
checkmate::assertIntegerish(iters, min.len = 1, any.missing = FALSE, null.ok = TRUE)
checkmate::assertCharacter(blearner.type, len = 1, null.ok = TRUE)
if (is.null(blearner.type)) {
stop("Please specify a valid base-learner plus feature.")
}
if (! blearner.type %in% names(private$bl.list)) {
stop("Your requested feature plus learner is not available. Check 'getBaselearnerNames()' for available learners.")
}
if (length(private$bl.list[[blearner.type]]$feature) > 1) {
stop("Only univariate plotting is supported.")
}
# Check if selected base-learner includes the proposed one + check if iters is big enough:
iter.min = which(self$getSelectedBaselearner() == blearner.type)[1]
if (! blearner.type %in% unique(self$getSelectedBaselearner())) {
stop("Requested base-learner plus feature was not selected.")
} else {
if (any(iters < iter.min)) {
warning("Requested base-learner plus feature was first selected at iteration ", iter.min)
}
}
feat.name = private$bl.list[[blearner.type]]$target$getIdentifier()
checkmate::assertNumeric(x = self$data[[feat.name]], min.len = 2, null.ok = FALSE)
checkmate::assertNumeric(from, lower = min(self$data[[feat.name]]), upper = max(self$data[[feat.name]]), len = 1, null.ok = TRUE)
checkmate::assertNumeric(to, lower = min(self$data[[feat.name]]), upper = max(self$data[[feat.name]]), len = 1, null.ok = TRUE)
if (is.null(from)) {
from = min(self$data[[feat.name]])
}
if (is.null(to)) {
to = max(self$data[[feat.name]])
}
if (from >= to) {
warning("Argument from is smaller than to, hence the x interval is [to, from].")
temp = from
from = to
to = temp
}
plot.data = as.matrix(seq(from = from, to = to, length.out = length.out))
feat.map = private$bl.list[[blearner.type]]$factory$transformData(plot.data)
# Create data.frame for plotting depending if iters is specified:
if (!is.null(iters[1])) {
preds = lapply(iters, function (x) {
if (x >= iter.min) {
return(feat.map %*% self$model$getParameterAtIteration(x)[[blearner.type]])
} else {
return(rep(0, length.out))
}
})
names(preds) = iters
df.plot = data.frame(
effect = unlist(preds),
iteration = as.factor(rep(iters, each = length.out)),
feature = plot.data
)
gg = ggplot2::ggplot(df.plot, ggplot2::aes(feature, effect, color = iteration))
} else {
df.plot = data.frame(
effect = feat.map %*% self$getEstimatedCoef()[[blearner.type]],
feature = plot.data
)
gg = ggplot2::ggplot(df.plot, ggplot2::aes(feature, effect))
}
# If there are too much rows we need to take just a sample or completely remove rugs:
if (nrow(self$data) > 1000) {
idx.rugs = sample(seq_len(nrow(self$data)), 1000, FALSE)
} else {
idx.rugs = seq_len(nrow(self$data))
}
gg = gg +
ggplot2::geom_line() +
ggplot2::geom_rug(data = self$data[idx.rugs,], ggplot2::aes_string(x = feat.name), inherit.aes = FALSE,
alpha = 0.8) +
ggplot2::xlab(feat.name) +
ggplot2::xlim(from, to) +
ggplot2::ylab("Additive Contribution") +
ggplot2::labs(title = paste0("Effect of ", blearner.type),
subtitle = "Additive contribution of predictor")
return(gg)
} else {
message("Please install ggplot2 to create plots.")
return(NULL)
}
},
getBaselearnerNames = function () {
# return(lapply(private$bl.list, function (bl) bl[[1]]$target$getIdentifier()))
return(names(private$bl.list))
}
),
private = list(
# Lists of single logger and base-learner factories. Necessary to prevent the factories from the
# arbage collector which deallocates all the data from the heap and couses R to crash.
l.list = list(),
bl.list = list(),
logger.list = list(),
initializeModel = function() {
private$logger.list = LoggerList$new()
for (n in names(private$l.list)) {
private$logger.list$registerLogger(n, private$l.list[[n]])
}
self$model = Compboost_internal$new(self$response, self$learning.rate,
self$stop.if.all.stoppers.fulfilled, self$bl.factory.list, self$loss, private$logger.list, self$optimizer)
},
addSingleNumericBl = function(data.columns, feature, id.fac, id, bl.factory, data.source, data.target, ...) {
private$bl.list[[id]] = list()
private$bl.list[[id]]$source = data.source$new(as.matrix(data.columns), paste(feature, collapse = "_"))
private$bl.list[[id]]$feature = feature
private$bl.list[[id]]$target = data.target$new()
# Call handler for default arguments and argument handling:
handler.name = paste0(".handle", bl.factory@.Data)
par.set = c(source = private$bl.list[[id]]$source, target = private$bl.list[[id]]$target, id = id.fac, do.call(handler.name, list(...)))
private$bl.list[[id]]$factory = do.call(bl.factory$new, par.set)
# private$bl.list[[id]]$factory = bl.factory$new(private$bl.list[[id]]$source, private$bl.list[[id]]$target, id.fac, ...)
self$bl.factory.list$registerFactory(private$bl.list[[id]]$factory)
private$bl.list[[id]]$source = NULL
},
addSingleCatBl = function(data.column, feature, id.fac, id, bl.factory, data.source, data.target, ...) {
lvls = unlist(unique(data.column))
# Create dummy variable for each category and use that vector as data matrix. Hence,
# if a categorical feature has 3 groups, then these 3 groups are added as 3 different
# base-learners (unbiased feature selection).
for (lvl in lvls) {
list.id = paste(feature, lvl, id.fac, sep = "_")
private$addSingleNumericBl(data.columns = as.matrix(as.integer(data.column == lvl)),
feature = paste(feature, lvl, sep = "_"), id.fac = id.fac,
id = list.id, bl.factory, data.source, data.target, ...)
# This is important because of:
# 1. feature in addSingleNumericBl needs to be something like cat_feature_Group1 to define the
# data objects correctly in a unique way.
# 2. The feature itself should not be named with the level. Instead of that we just want the
# feature name of the categorical variable, such as cat_feature (important for predictions).
private$bl.list[[list.id]]$feature = feature
}
}
)
)
Try the compboost package in your browser
Any scripts or data that you put into this service are public.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.