PRTools manual in preparation -- Dataset, creation of artificial datasets

Dataset, creation of artificial datasets

Datasets can be created in the following ways:

They may already exist as Matlab mat-files. Data collected in the past and converted to PRTools datasets can be stored by the Matlab save command and loaded by the load command. Users are encouraged to build a function for doing this as the help comments in the function may be used to give additional information. Alternatively or inside such a routine the PRTools command prdataset may be used to load and handle the mat-file.

A set of public domain datasets in PRTools format can be downloaded from the PRTools web site.

From the analysis of objects like images and time signals features may be extracted. The resulting matrix of feature vectors a dataset can be defined, see dataset definition. An easy way to define the class labels is offered by genlab. Datafiles may be used to define the preprocessing on data items stored as files on disk. The section on image classification gives an example.

For research on the design and evaluation of classifiers the generation of artificial datasets can be very useful as they enable to define exactly the data model and to generate an arbitrary amount of data. Random generators like the Matlab rand and randn commands (in combination with genlab) and the PRTools gauss routine may be used for this.

PRTools offers a set of commands for generating artificial datasets, see below.

> Commands for generating artificial datasets
`gendatb`	Generation of two banana shaped classes.
`gendatc`	Generation of two circular classes.
`gendatd`	Generation of two normally distributed 'difficult' classes. This example is 'difficult' as the distribution should be known to construct a good classifier in case of small sample sizes.
`gendath`	Generation of two normally distributed classes according to Highleyman [], see also [].
`gendatl`	Generation of the 'Lithuanian' classes as proposed by Raudys.
`gendats`	Generation of two simple normally distributed classes.
`gendatm`	Generation of eight 2d classes.
`gentrunk`	Generation of Trunk's dataset [], used to illustrate the peaking phenomenon (curse of dimensionality).
`gendatgauss`	Generation of multivariate Gaussian distributed data.
`gencirc`	Generation of a one-class circular dataset.
`circles3d`	Create a dataset containing 2 circles in 3 dimensions (for mds examples).
`lines5d`	Create a dataset containing 3 lines in 5 dimensions (for mds examples).
`gendatr`	Generate regression dataset from data and target values (for regression examples).
`gendat`	Random sampling of datasets for training and testing.
`gensubsets`	Generation of a series of consistent subsets of a dataset.
`gendatk`	Nearest neighbor data generation.
`gendatp`	Parzen density data generation.

The last three commands generate datasets out of existing datasets.

R.P.W. Duin, January 28, 2013

Dataset download of real world datasets: prdatasets

Dataset, creation of artificial datasets