Revision 401d48b8bef1ac9992b5a017291778be931d0d2a authored by Patrick Schratz on 16 April 2020, 14:06:11 UTC, committed by GitHub on 16 April 2020, 14:06:11 UTC
1 parent 0cb2918
usecase_regression.Rmd
---
title: "Use case: Regression"
output: rmarkdown::html_vignette
vignette: >
%\VignetteIndexEntry{mlr}
%\VignetteEngine{knitr::rmarkdown}
\usepackage[utf8]{inputenc}
---
```{r, echo = FALSE, message=FALSE}
library("mlr")
library("BBmisc")
library("ParamHelpers")
# show grouped code output instead of single lines
knitr::opts_chunk$set(collapse = TRUE)
```
For the regression use case we use the well-known `mlbench::BostonHousing()` dataset.
The road map is as follows:
* define the learning task ([here](task.html)),
* tune the model ([here](tune.html)),
* conduct a benchmark experiment ([here](benchmark_experiments.html)) and
* evaluate the performance of the model ([here](performance.html)).
First, let's have a look at the data.
```{r}
data(BostonHousing, package = "mlbench")
summary(BostonHousing)
```
This data set concerns housing in the suburban area of Boston.
The target variable, chosen for the regression task, is `medv` - the median value of owner-occupied homes in \$1000's.
Description of the others 13 attributes can be found at `mlbench::BostonHousing()`).
# Define a task
Now, let us continue with defining the regression [task](task.html).
```{r}
# Make a task
regr.task = makeRegrTask(data = BostonHousing, target = "medv")
regr.task
```
In order to get an overview of the features type, we can print out the `regr.task`.
This shows that there are 12 numeric and one factor variables in the data set.
# Tuning
By calling `listLearners("regr")` we can see which learners are available for the regression task.
With so many learners it is difficult to choose which one would be optimal for this specific task.
As such we will choose a sample of these learners and compare their results.
This analysis uses the classical linear regression model (`regr.lm`), SVM (`kernlab::ksvm()`) with a radial basis kernel (`regr.ksvm`) and random forest from the ranger (`ranger::ranger()`) package (`regr.ranger`).
In order to get a quick overview of all learner-specific tunable parameters you can call `getLearnerParamSet()` or its alias `ParamHelpers::getParamSet()`, which will list learner's hyperparameters and their properties.
Before setting up a [benchmark experiment](benchmark_experiments.html) we can specify which hyperparameters are going to be tuned.
The `mlr` package provides powerful tuning algorithms, such as iterated F-racing (`irace::irace()`), CMA Evolution Strategy (`cmaes::cma_es()`), model-based / Bayesian optimization (`mlrMBO::mbo()`) and generalized simulated annealing (`GenSA::GenSA()`).
See [Tuning](tuning.html) and [Advanced Tuning](advanced_tune.html) for more details.
For each learner one hyperparameter will be tuned, i.e. kernel parameter `sigma` for SVM model and the number of trees (`num.trees`) in the random forest model.
We start with specifying a search space for each of these parameters.
With `makeTuneControlCMAES()` we set the tuning method to be CMA Evolution Strategy (`cmaes::cma_es()`).
Afterwards we take 5-fold cross validation as our [resampling strategy](resample.html) and root mean squared error (`rmse`) as optimization criterion.
Finally, we make tuning wrapper for each learner.
```{r}
set.seed(1234)
# Define a search space for each learner'S parameter
ps_ksvm = makeParamSet(
makeNumericParam("sigma", lower = -12, upper = 12, trafo = function(x) 2^x)
)
ps_rf = makeParamSet(
makeIntegerParam("num.trees", lower = 1L, upper = 200L)
)
# Choose a resampling strategy
rdesc = makeResampleDesc("CV", iters = 5L)
# Choose a performance measure
meas = rmse
# Choose a tuning method
ctrl = makeTuneControlCMAES(budget = 100L)
# Make tuning wrappers
tuned.ksvm = makeTuneWrapper(learner = "regr.ksvm", resampling = rdesc, measures = meas,
par.set = ps_ksvm, control = ctrl, show.info = FALSE)
tuned.rf = makeTuneWrapper(learner = "regr.ranger", resampling = rdesc, measures = meas,
par.set = ps_rf, control = ctrl, show.info = FALSE)
```
# Benchmark Experiment
In order to conduct a [benchmark experiment](benchmark_experiments.html), it is necessary to choose an evaluation method.
We will use the resampling strategy and the performance measure from the previous section and then pass the tuning wrappers as arguments into the `benchmark()` function.
```{r}
# Four learners to be compared
lrns = list(makeLearner("regr.lm"), tuned.ksvm, tuned.rf)
# Conduct the benchmark experiment
bmr = benchmark(learners = lrns, tasks = regr.task, resamplings = rdesc, measures = rmse,
show.info = FALSE)
```
# Performance
Now we want to evaluate the results.
```{r}
getBMRAggrPerformances(bmr)
```
A closer look at the boxplot reveals that RF outperforms the other learners for this specific task.
Despite the tuning procedure performed before, the benchmark experiment for linear and lasso regression yields similar but poor results.
```{r}
plotBMRBoxplots(bmr)
```
![swh spinner](/static/img/swh-spinner.gif)
Computing file changes ...