spFeatureSelection: SPSA-FSR for Feature Selection and Ranking

Description

This function searches for the best performing features and rank the feature importance by implementing simultaneous perturbation stochastic approximation (SPSA) algorithm given a task and a wrapper. The task and wrapper are defined using the mlr3 package.

Usage

spFeatureSelection(task, wrapper = NULL, scoring = NULL, ...)

Value

spFSR returns an object of class "spFSR". An object of class "spFSR" consists of the following:

task.spfs: An mlr3 package tsk object defined on the best performing features.
wrapper: An mlr3 package lrn object, default is random forest.
scoring: An mlr3 package msr object as specified by the user.
param best.model: An mlr3 package model object trained by the wrapper using task.spfs.
iter.results: A data.frame object containing detailed information on each iteration.
features: Names of the best performing features.
num.features: The number of best performing features.
importance: A vector of importance ranks of the best performing features.
total.iters: The total number of iterations executed.
best.iter: The iteration where the best performing feature subset was encountered.
best.value: The best measure value encountered during execution.
best.std: The standard deviation corresponding to the best measure value encountered.
run.time: Total run time in minutes
results: Dataframe with boolean of selected features, names and measure
call: Call

Arguments

task: A task object created using mlr3 package. It must be either a ClassifTask or RegrTask object.
wrapper: A Learner object created using mlr3 package. Multiple learners object is not supported.
scoring: A performance measure within the mlr3 package supported by the task.
...: Additional arguments. For more details, see spFSR.default.

References

David V. Akman et al. (2022) k-best feature selection and ranking via stochastic approximation, Expert Systems with Applications, Vol. 213. See tools:::Rd_expr_doi("10.1016/j.eswa.2022.118864")

G.F.A Yeo and V. Aksakalli (2021) A stochastic approximation approach to simultaneous feature weighting and selection for nearest neighbour learners, Expert Systems with Applications, Vol. 185. See tools:::Rd_expr_doi("10.1016/j.eswa.2021.115671")

Examples

Run this code

library(mlr3)          # load the mlr3 package
library(mlr3learners) # load the mlr3learners package

task    <- tsk('iris') # define task
wrapper <- lrn('classif.rpart')                # define wrapper
measure <- msr('classif.acc')

# run spsa
spsaMod <- spFeatureSelection( task = task,
                               wrapper = wrapper,
                               scoring = measure,
                               num.features.selected = 3,
                               n.jobs = 1,
                               iters.max = 2,
                               num.grad.avg = 1)


# obtain summary
summary(spsaMod)

# plot spsaMod
plot(spsaMod)                                # simplest plot
plot(spsaMod, errorBar = TRUE)               # plot with error bars
plot(spsaMod, errorBar = TRUE, se = TRUE)    # plot with error bars based on se
plot(spsaMod, errorBar = TRUE, annotateBest = TRUE)  # annotate best value
plot(spsaMod, errorBar = TRUE, ylab = 'Acc measure', type = 'o')

# obtain the wrapped model with the best performing features
bestMod <- getBestModel(spsaMod)

# predict using the best mod
pred <- bestMod$predict( task = spsaMod$task.spfs )

# Obtain confusion matrix
pred$confusion

# Get the importance ranks of best performing features
getImportance(spsaMod)
plotImportance(spsaMod)