Branch: master
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
115 lines (88 sloc) 4.18 KB
title: "Generic Bagging"
output: rmarkdown::html_vignette
vignette: >
```{r, echo = FALSE, message=FALSE}
# show grouped code output instead of single lines
knitr::opts_chunk$set(collapse = TRUE)
One reason why random forests perform so well is that they are using bagging as
a technique to gain more stability. But why do you want to limit yourself to the
classifiers already implemented in well known random forests when it is really easy to build your own with `mlr`?
Just bag an `mlr` learner already `makeBaggingWrapper()`.
As in a random forest, we need a Learner which is trained on a subset of the data during each iteration of the bagging process.
The subsets are chosen according to the parameters given to `makeBaggingWrapper()`:
* `bw.iters` On how many subsets (samples) do we want to train our Learner?
* `bw.replace` Sample with replacement (also known as *bootstrapping*)?
* `bw.size` Percentage size of the samples. If `bw.replace = TRUE`, `bw.size = 1` is the default. This does not mean that one sample will contain all the observations as observations will occur multiple times in each sample.
* `bw.feats` Percentage size of randomly selected features for each iteration.
Of course we also need a Learner which we have to pass to `makeBaggingWrapper()`.
```{r makeBaggingWrapper_setup}
lrn = makeLearner("classif.rpart")
bag.lrn = makeBaggingWrapper(lrn, bw.iters = 50, bw.replace = TRUE,
bw.size = 0.8, bw.feats = 3/4)
Now we can compare the performance with and without bagging.
First let's try it without bagging:
```{r makeBaggingWrapper_without}
rdesc = makeResampleDesc("CV", iters = 10)
r = resample(learner = lrn, task = sonar.task, resampling = rdesc, = FALSE)
And now with bagging:
```{r makeBaggingWrapper_with}
rdesc = makeResampleDesc("CV", iters = 10)
result = resample(learner = bag.lrn, task = sonar.task, resampling = rdesc, = FALSE)
Training more learners takes more time, but can outperform pure learners
on noisy data with many features.
# Changing the type of prediction
In case of a *classification* problem the predicted class labels are determined by majority voting over the predictions of the individual models.
Additionally, posterior probabilities can be estimated as the relative proportions of the predicted class labels.
For this purpose you have to change the predict type of the *bagging learner* as follows.
```{r, eval = FALSE}
bag.lrn = setPredictType(bag.lrn, predict.type = "prob")
Note that it is not relevant if the *base learner* itself can predict probabilities and that for this reason the predict type of the *base learner* always has to be ``"response"``.
For *regression* the mean value across predictions is computed.
Moreover, the standard deviation across predictions is estimated if the predict type of the bagging learner is changed to ``"se"``.
Below, we give a small example for regression.
```{r makeBaggingWrapper_regression}
n = getTaskSize(bh.task)
train.inds = seq(1, n, 3)
test.inds = setdiff(1:n, train.inds)
lrn = makeLearner("regr.rpart")
bag.lrn = makeBaggingWrapper(lrn)
bag.lrn = setPredictType(bag.lrn, predict.type = "se")
mod = train(learner = bag.lrn, task = bh.task, subset = train.inds)
With function `getLearnerModel()`, you can access the models fitted in the
individual iterations.
head(getLearnerModel(mod), 2)
Predict the response and calculate the standard deviation:
pred = predict(mod, task = bh.task, subset = test.inds)
In the column labelled ``se`` the standard deviation for each prediction is given.
Let's visualise this a bit using `ggplot2::ggplot2()`.
Here we plot the percentage of lower status of the population (`lstat`) against the prediction.
```{r makeBaggingWrapper_regressionPlot}
data = cbind(, getTaskData(bh.task, subset = test.inds))
g = ggplot(data, aes(x = lstat, y = response, ymin = response - se,
ymax = response + se, col = age))
g + geom_point() + geom_linerange(alpha = 0.5)