tune.spLearner-methods: Optimize spLearner by fine-tuning parameters and running...

In Envirometrix/landmap: Automated Spatial Prediction using Ensemble Machine Learning

Optimize spLearner by fine-tuning parameters and running feature selection

Description

Optimize spLearner by fine-tuning parameters and running feature selection

Usage

## S4 method for signature 'spLearner'
tune.spLearner(
  object,
  num.trees = 85,
  blocking,
  discrete_ps,
  rdesc = mlr::makeResampleDesc("CV", iters = 2L),
  inner = mlr::makeResampleDesc("Holdout"),
  maxit = 20,
  xg.model_Params,
  xg.skip = FALSE,
  parallel = "multicore",
  hzn_depth = FALSE,
  ...
)

Arguments

object	spLearner object (unoptimized),
num.trees	number of random forest trees,
blocking	blocking columns,
discrete_ps	settings for random forest,
rdesc	resampling method for fine-tuning,
inner	resampling method for feature selection,
maxit	maximum number of iterations for feature selection,
xg.model_Params	xgboost parameter set,
xg.skip	logical, should the tuning of the XGboost should be skipped?
parallel	Initiate parallel processing,
hzn_depth	specify whether horizon depth available in the training dataframe,
...	other arguments that can be passed on to mlr::makeStackedLearner,

Value

optimized object of type spLearner

Note

Currently requires that two base learners are regr.ranger and regr.xgboost, and that there are at least 3 base learners in total. Fine-tuning and feature selection can be quite computational and it is highly recommended to start with smaller subsets of data and then measure processing time. The function mlr::makeFeatSelWrapper can result in errors if the covariates have a low variance or follow a zero-inflated distribution. Reducing the number of features via feature selection and fine-tuning of the Random Forest mtry and XGboost parameters, however, can result in significantly higher prediction speed and accuracy.

Author(s)

Tom Hengl

Examples

library(mlr)
library(ParamHelpers)
library(xgboost)
library(kernlab)
library(ranger)
library(glmnet)
library(raster)
demo(meuse, echo=FALSE)
## Regression:
sl = c("regr.ranger", "regr.xgboost", "regr.ksvm", "regr.cvglmnet")
m <- train.spLearner(meuse["lead"], covariates=meuse.grid[,c("dist","ffreq")],
      lambda=0, parallel=FALSE, SL.library=sl)
summary(m@spModel$learner.model$super.model$learner.model)
## Optimize model:
t <- try( m0 <- tune.spLearner(m, xg.skip = TRUE, parallel=FALSE), silent=TRUE)
if(!class(t) == "try-error") summary(m0@spModel$learner.model$super.model$learner.model)

Envirometrix/landmap documentation built on June 10, 2022, 10:12 p.m.