fit_stacked_model: Fit a stacking model that assigns weights to component models...

In reichlab/adaptively-weighted-ensemble: Adaptively Weighted Ensembles via Stacking

Description

Fit a stacking model that assigns weights to component models The weights are a function of observed covariates (which?), and are obtained via gradient tree boosting

Usage

fit_stacked_model(regions, prediction_target, component_model_names,
  explanatory_variables = c("analysis_time_season_week",
  "kcde_model_confidence", "sarima_model_confidence", "weighted_ili"),
  loso_preds_path, seasons_to_leave_out, booster = "gbtree", subsample = 1,
  colsample_bytree = 1, colsample_bylevel = 1, max_depth = 10,
  min_child_weight = -10^10, eta = 0.3, gamma = 0, lambda = 0,
  alpha = 0, nrounds = 10, cv_params = NULL, cv_folds = NULL,
  cv_nfolds = 10L, cv_refit = "ttest", update = NULL, nthread = NULL,
  verbose = 0)

Arguments

regions	string with region: either "National" or in the format "Regionk" for k in 1, ..., 10
prediction_target	string with either "onset", "peak_week", "peak_inc", "ph1_inc", ..., "ph4_inc"
component_model_names	character vector with names of component models
explanatory_variables	character vector with names of explanatory variables to include for weights; a non-empty subset of "analysis_time_season_week", "kcde_model_confidence", "sarima_model_confidence", "weighted_ili"
loso_preds_path	path to directory with leave-one-season-out predictions from each component model. Predictions should be saved in files named like "kde-National-loso-predictions.rds"
seasons_to_leave_out	optional character vector of seasons to leave out of stacking estimation
booster	what form of boosting to use? see xgboost documentation
subsample	fraction of data to use in bagging. not supported yet.
colsample_bytree	fraction of explanatory variables to randomly select in growing each regression tree. see xgboost documentation
colsample_bylevel	fraction of explanatory variables to randomly select in growing each level of the regression tree. see xgboost documentation
max_depth	maximum depth of regression trees. see xgboost documentation
min_child_weight	not recommended for use. see xgboost documentation
eta	learning rate. see xgboost documentation
gamma	Penalty on number of regression tree leafs. see xgboost documentation
lambda	L2 regularization of contribution to model weights in each round. see xgboost documentation
alpha	L1 regularization of contribution to model weights in each round. see xgboost documentation
nrounds	see xgboost documentation
cv_params	optional named list of parameter values to evaluate loss via cross-validation. Each component is a vector of parameter values with name one of "booster", "subsample", "colsample_bytree", "colsample_bylevel", "max_depth", "min_child_weight", "eta", "gamma", "lambda", "alpha", "nrounds"
cv_folds	list specifying observation groups to use in cross-validation each list component is a numeric vector of observation indices.
cv_nfolds	integer specifying the number of cross-validation folds to use. if cv_folds was provided, cv_nfolds is ignored. if cv_folds was not provided, the data will be randomly partitioned into cv_nfolds groups
cv_refit	character describing which of the models specified by the values in cv_params to refit using the full data set. Either "best", "ttest", or "none".
update	an object of class xgbstack to update
nthread	how many threads to use. see xgboost documentation
verbose	how much output to generate along the way. 0 for no logging, 1 for some logging

Value

a model stacking fit

reichlab/adaptively-weighted-ensemble documentation built on May 27, 2019, 4:53 a.m.