Skip to content
Permalink
Branch: master
Find file Copy path
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
251 lines (176 sloc) 10.8 KB
---
title: "Learning Tasks"
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)
set.seed(123)
```
Learning tasks encapsulate the data set and further relevant information about a machine learning problem, for example the name of the target variable for supervised problems.
# Task types and creation
The tasks are organized in a hierarchy, with the generic `Task()` at the top.
The following tasks can be instantiated and all inherit from the virtual superclass `Task()`:
* `RegrTask()` for regression problems,
* `ClassifTask()` for binary and [multi-class classification problems](cost_sensitive_classif.html){target="_blank"} with class-dependent costs can be handled as well),
* `SurvTask()` for survival analysis,
* `ClusterTask()` for cluster analysis,
* `MultilabelTask()` for multilabel classification problems,
* `CostSensTask()` for general [cost sensitive classification](cost_sensitive_classif.html){target="_blank"} (with example-specific costs).
To create a task, just call ``make<TaskType>``, e.g., `makeClassifTask()`.
All tasks require an identifier (argument ``id``) and a `base::data.frame()` (argument ``data``).
If no ID is provided it is automatically generated using the variable name of the data.
The ID will be later used to name results, for example of [benchmark experiments](benchmark_experiments.html){target="_blank"}, and to annotate plots.
Depending on the nature of the learning problem, additional arguments may be required and are discussed in the following sections.
## Regression
For supervised learning like regression (as well as classification and survival analysis) we, in addition to ``data``, have to specify the name of the ``target`` variable.
```{r}
data(BostonHousing, package = "mlbench")
regr.task = makeRegrTask(id = "bh", data = BostonHousing, target = "medv")
regr.task
```
As you can see, the `Task()` records the type of the learning problem and basic information about the data set, e.g., the types of the features (`base::numeric()` vectors, `base::factors()` or ordered factors), the number of observations, or whether missing values are present.
Creating tasks for classification and survival analysis follows the same scheme, the data type of the target variables included in ``data`` is simply different.
For each of these learning problems some specifics are described below.
## Classification
For classification the target column has to be a `factor`.
In the following example we define a classification task for the `mlbench::BreastCancer()` data set and exclude the variable ``Id`` from all further model fitting and evaluation.
```{r}
data(BreastCancer, package = "mlbench")
df = BreastCancer
df$Id = NULL
classif.task = makeClassifTask(id = "BreastCancer", data = df, target = "Class")
classif.task
```
In binary classification the two classes are usually referred to as *positive* and *negative* class with the positive class being the category of greater interest.
This is relevant for many [performance measures](performance.html){target="_blank"} like the *true positive rate* or [ROC analysis](roc_analysis.html){target="_blank"}.
Moreover, `mlr`, where possible, permits to set options (like the `setThreshold()` or `makeWeightedClassesWrapper()`) and returns and plots results (like class posterior probabilities) for the positive class only.
`makeClassifTask()` by default selects the first factor level of the target variable as the positive class, in the above example ``benign``.
Class ``malignant`` can be manually selected as follows:
```{r}
classif.task = makeClassifTask(id = "BreastCancer", data = df, target = "Class", positive = "malignant")
```
## Survival analysis
Survival tasks use two target columns.
For left and right censored problems these consist of the survival time and a binary event indicator.
For interval censored data the two target columns must be specified in the ``"interval2"`` format (see `survival::Surv()`).
```{r}
data(lung, package = "survival")
lung$status = (lung$status == 2) # convert to logical
surv.task = makeSurvTask(data = lung, target = c("time", "status"))
surv.task
```
The type of censoring can be specified via the argument ``censoring``, which defaults to ``"rcens"`` for right censored data.
## Multilabel classification
In multilabel classification each object can belong to more than one category at the same time.
The `data` are expected to contain as many target columns as there are class labels.
The target columns should be logical vectors that indicate which class labels are present.
The names of the target columns are taken as class labels and need to be passed to the `target` argument of `makeMultilabelTask()`.
In the following example we get the data of the yeast data set, extract the label names, and pass them to the ``target`` argument in `makeMultilabelTask()`.
```{r}
yeast = getTaskData(yeast.task)
labels = colnames(yeast)[1:14]
yeast.task = makeMultilabelTask(id = "multi", data = yeast, target = labels)
yeast.task
```
See also the tutorial page [multilabel](multilabel.html){target="_blank"}.
## Cluster analysis
As cluster analysis is unsupervised, the only mandatory argument to construct a cluster analysis task is the ``data``.
Below we create a learning task from the data set `datasets::mtcars()`.
```{r}
data(mtcars, package = "datasets")
cluster.task = makeClusterTask(data = mtcars)
cluster.task
```
## Cost-sensitive classification
The standard objective in classification is to obtain a high prediction accuracy, i.e., to minimize the number of errors.
All types of misclassification errors are thereby deemed equally severe.
However, in many applications different kinds of errors cause different costs.
In case of *class-dependent costs*, that solely depend on the actual and predicted class labels, it is sufficient to create an ordinary `ClassifTask()`.
In order to handle *example-specific costs* it is necessary to generate a `CostSensTask()`.
In this scenario, each example $(x, y)$ is associated with an individual cost vector of length $K$ with $K$ denoting the number of classes.
The $k$-th component indicates the cost of assigning $x$ to class $k$. Naturally, it is assumed that the cost of the intended class label $y$ is minimal.
As the cost vector contains all relevant information about the intended class $y$, only the feature values $x$ and a ``cost`` matrix, which contains the cost vectors for all examples in the data set, are required to create the `CostSensTask()`.
In the following example we use the `datasets::iris()` data and an artificial cost matrix (which is generated as proposed by [Beygelzimer et al., 2005](https://doi.org/10.1145/1102351.1102358)):
```{r}
df = iris
cost = matrix(runif(150 * 3, 0, 2000), 150) * (1 - diag(3))[df$Species,]
df$Species = NULL
costsens.task = makeCostSensTask(data = df, cost = cost)
costsens.task
```
For more details see the page on [cost sensitive classification](cost_sensitive_classif.html){target="_blank"}.
# Further settings
The `Task()` help page also lists several other arguments to describe further details of the learning problem.
For example, we could include a ``blocking`` factor in the task.
This would indicate that some observations "belong together" and should not be separated when splitting the data into training and test sets for [resampling](resample.html){target="_blank"}.
Another option is to assign ``weights`` to observations.
These can simply indicate observation frequencies or result from the sampling scheme used to collect the data.
Note that you should use this option only if the weights really belong to the task.
If you plan to train some learning algorithms with different weights on the same `Task()`, `mlr` offers several other ways to set observation or class weights (for supervised classification).
See for example the tutorial page about [training](train.html){target="_blank"} or function `makeWeightedClassesWrapper()`.
# Accessing a learning task
We provide many operators to access the elements stored in a `Task()`.
The most important ones are listed in the documentation of `Task()` and `getTaskData()`.
To access the `TaskDesc()` that contains basic information about the task you can use:
```{r}
getTaskDesc(classif.task)
```
Note that `TaskDesc()` have slightly different elements for different types of `Task()`s.
Frequently required elements can also be accessed directly.
```{r}
# Get the ID
getTaskId(classif.task)
# Get the type of task
getTaskType(classif.task)
# Get the names of the target columns
getTaskTargetNames(classif.task)
# Get the number of observations
getTaskSize(classif.task)
# Get the number of input variables
getTaskNFeats(classif.task)
# Get the class levels in classif.task
getTaskClassLevels(classif.task)
```
Moreover, `mlr` provides several functions to extract data from a `Task()`.
```{r}
# Accessing the data set in classif.task
str(getTaskData(classif.task))
# Get the names of the input variables in cluster.task
getTaskFeatureNames(cluster.task)
# Get the values of the target variables in surv.task
head(getTaskTargets(surv.task))
# Get the cost matrix in costsens.task
head(getTaskCosts(costsens.task))
```
Note that `getTaskData()` offers many options for converting the data set into a convenient format.
This especially comes in handy when you [integrate a new learner](create_learner.html){target="_blank"} from another **R** package into `mlr`.
In this regard function `getTaskFormula()` is also useful.
# Modifying a learning task
`mlr` provides several functions to alter an existing `Task()`, which is often more convenient than creating a new `Task()` from scratch.
Here are some examples.
```{r}
# Select observations and/or features
cluster.task = subsetTask(cluster.task, subset = 4:17)
# It may happen, especially after selecting observations, that features are constant.
# These should be removed.
removeConstantFeatures(cluster.task)
# Remove selected features
dropFeatures(surv.task, c("meal.cal", "wt.loss"))
# Standardize numerical features
task = normalizeFeatures(cluster.task, method = "range")
summary(getTaskData(task))
```
For more functions and more detailed explanations have a look at the [data preprocessing](preproc.html){target="_blank"} page.
# Example tasks and convenience functions
For your convenience `mlr` provides pre-defined `Task()`s for each type of learning problem.
These are also used throughout this tutorial in order to get shorter and more readable code.
A list of all `Task()`s can be found in the [Appendix](example_tasks.html){target="_blank"}.
Moreover, `mlr`'s function `convertMLBenchObjToTask()` can generate `Task()`s from the data sets and data generating functions in package `mlbench::mlbench()`.
You can’t perform that action at this time.