CAST (version 0.4.1)

ffs: Forward feature selection

Description

A simple forward feature selection algorithm

Usage

ffs(predictors, response, method = "rf",
  metric = ifelse(is.factor(response), "Accuracy", "RMSE"),
  maximize = ifelse(metric == "RMSE", FALSE, TRUE), withinSE = FALSE,
  minVar = 2, trControl = caret::trainControl(), tuneLength = 3,
  tuneGrid = NULL, seed = sample(1:1000, 1), verbose = TRUE, ...)

Arguments

predictors

see train

response

see train

method

see train

metric

see train

maximize

see train

withinSE

Logical Models are only selected if they are better than the currently best models Standard error

minVar

Numeric. Number of variables to combine for the first selection. See Details.

trControl

see train

tuneLength

see train

tuneGrid

see train

seed

A random number used for model training

verbose

Logical. Should information about the progress be printed?

...

arguments passed to the classification or regression routine (such as randomForest).

Value

A list of class train. Beside of the usual train content the object contains the vector "selectedvars" and "selectedvars_perf" that give the order of the best variables selected as well as their corresponding performance (starting from the first two variables). It also contains "perf_all" that gives the performance of all model runs.

Details

Models with two predictors are first trained using all possible pairs of predictor variables. The best model of these initial models is kept. On the basis of this best model the predictor variables are iteratively increased and each of the remaining variables is tested for its improvement of the currently best model. The process stops if none of the remaining variables increases the model performance when added to the current best model.

The internal cross validation can be run in parallel. See information on parallel processing of carets train functions for details.

Using withinSE will favour models with less variables and probably shorten the calculation time

Per Default, the ffs starts with all possible 2-pair combinations. minVar allows to start the selection with more than 2 variables, e.g. minVar=3 starts the ffs testing all combinations of 3 (instead of 2) variables first and then increasing the number. This is important for e.g. neural networks that often cannot make sense of only two variables. It is also relevant if it is assumed that the optimal variables can only be found if more than 2 are considered at the same time.

References

• Gasch, C.K., Hengl, T., Gr<U+00E4>ler, B., Meyer, H., Magney, T., Brown, D.J. (2015): Spatio-temporal interpolation of soil water, temperature, and electrical conductivity in 3D+T: the Cook Agronomy Farm data set. Spatial Statistics 14: 70-90.

• Meyer, H., Reudenbach, C., Hengl, T., Katurji, M., Nau<U+00DF>, T. (2018): Improving performance of spatio-temporal machine learning models using forward feature selection and target-oriented validation. Environmental Modelling & Software 101: 1-9.

• Meyer, H., Reudenbach, C., W<U+00F6>llauer, S., Nauss, T. (2019): Importance of spatial predictor variable selection in machine learning applications - Moving from data reproduction to spatial prediction. Ecological Modelling. 411, 108815.

See Also

train,bss, trainControl,CreateSpacetimeFolds

Examples

# NOT RUN {
data(iris)
ffsmodel <- ffs(iris[,1:4],iris$Species)
ffsmodel$selectedvars
ffsmodel$selectedvars_perf
# }
# NOT RUN {
# or perform model with target-oriented validation (LLO CV)
#the example is taken from the GSIF package and is described
#in Gasch et al. (2015). The ffs approach for this dataset is described in
#Meyer et al. (2018). Due to high computation time needed, only a small and thus not robust example
#is shown here.

# }
# NOT RUN {
#run the model on three cores:
library(doParallel)
cl <- makeCluster(3)
registerDoParallel(cl)

#load and prepare dataset:
dat <- get(load(system.file("extdata","Cookfarm.RData",package="CAST")))
trainDat <- dat[dat$altitude==-0.3&year(dat$Date)==2012&week(dat$Date)%in%c(13:14),]

#visualize dataset:
ggplot(data = trainDat, aes(x=Date, y=VW)) + geom_line(aes(colour=SOURCEID))

#create folds for Leave Location Out Cross Validation:
set.seed(10)
indices <- CreateSpacetimeFolds(trainDat,spacevar = "SOURCEID",k=3)
ctrl <- trainControl(method="cv",index = indices$index)

#define potential predictors:
predictors <- c("DEM","TWI","BLD","Precip_cum","cday","MaxT_wrcc",
"Precip_wrcc","NDRE.M","Bt","MinT_wrcc","Northing","Easting")

#run ffs model with Leave Location out CV
set.seed(10)
ffsmodel <- ffs(trainDat[,predictors],trainDat$VW,method="rf",
tuneLength=1,trControl=ctrl)
ffsmodel

#compare to model without ffs:
model <- train(trainDat[,predictors],trainDat$VW,method="rf",
tuneLength=1, trControl=ctrl)
model
stopCluster(cl)
# }