makeCorrelationImageProblem: Correlation Image Problem

In ja-thomas/dlbench: Deep Learning Benchmark Problems

Description

This benchmark is motivated by http://guessthecorrelation.com/. The task is to assess the correlation between two variables visually by a scatterplot of sampled points. Each observation is generated as follows:

• A random correlation value is drawn uniformly between -1 and 1.

• points.per.image points are sampled from a bivariate normal distribution with zero mean, sampled correlation and a marginal variance of 1.

• The observations are scaled between 0 and 1.

• An image of the data is generated and saved in a given resolution and pointsize.

• The values of the pixels are returned as a vector with the correlation as target.

Usage

makeCorrelationImageProblem(n = 5000L, resolution = c(32, 32),
  points.per.image = 100, pointsize = 2, create.task = TRUE)

Arguments

n	[integer(1)] Number of observations to create. Default is 5000L.
resolution	[integer(2)] Resolution of the image. This defines the number of features. Default is c(32, 32), which creates 1024 features.
points.per.image	[integer(1)] How man points should be generated per image. Default is 100.
pointsize	[integer(1)] Pointsize of intepreted big points (1/72 inch). See bmp. Default is 2.
create.task	[logical(1)] Should a task be created? If TRUE a mlr regression task (see makeRegrTask is returned. Otherwise a data.frame is returned.

Value

RegrTask | data.frame.

Examples

 data = makeCorrelationImageProblem(n = 20, create.task = FALSE)
 image(matrix(as.numeric(data[1, -1]), ncol = 32))

ja-thomas/dlbench documentation built on May 18, 2019, 7:15 a.m.