learning.learn: Apply a Set of Machine Learning Methods and Produce a Final...

In thomasWeise/learnerSelectoR: An R Package for Automatically Selecting and Applying a (Machine) Learning Method

Description

This method applies a set of machine learning methods to a machine learning task. It does so by first performing cross-valdation to choose the method which most likely can generalize well, and then applies this method to the complete dataset.

It supports using different data representations.

The learning effort value q tells the system how much effort to invest into the learning. The value is handed down to the learners which may interpret it in a meaningful way. q=0 is the smallest possible effort value and means "Learn as fast as possible, don't care about the result quality." q=1 is the largest possible effort value, meaning "Do whatever you can to increase the learning quality, don't care about the runtime." During the cross-validation, phase the learners are applied with q^3, which should speed-up this selection procedure, while maintaining that q=1 would still excert maximum effort. During the final application of the selected learner, q is supplied as-is.

Usage

learning.learn(data, data.size, learners, test.quality,
  selector = .def.selector, representations = NULL,
  test.selector = selector, q = 0.75, threshold = 0.004)

Arguments

data	the data set based on which we perform the learning task
data.size	the number of elements in the data, i.e., how many samples we can use, which will be the basis for cross-validation
learners	a list of learners, functions which accept the output of the selectors and return an instance of learning.Result
test.quality	the quality metric used to get the solution quality on the test data
selector	a function which returns a sub-set of a data representation made suitable for learning: it takes as input a data representation and an integer array with the selected items and returns a corresponding subset. If the second argument is NULL, the selector should return the complete represented data. Via the selector function, it is also possible to cache quality measures or something.
representations	a vector or list of data representations which should be used for learning, or NULL to only use the original data.
test.selector	similar to selector, but used to derive the data for testing, by default equal to selector
q	the effort to spent in learning, a value between 0 (min) and 1 (max). Higher values may lead to much more computational time, lower values to potentially lower result quality.
threshold	the relative difference between two test qualities below which we will pick the the "smaller" result

Value

the result of the learning process

thomasWeise/learnerSelectoR documentation built on May 16, 2019, 7:47 a.m.