Data generator with a control over linear separability

We propose a new data generator that is useful for a systematic benchmarking of algorithms for classification and clustering.

Features

• A user can adjust:
  • how many pairs of classes must be linearly non-separable
  • the number of classes
  • the number of data-points inside a class
  • the probability distribution of data-points
  • the minimal distance between each pair of classes
  • the shape of a point-set that forms a class
• 38 different shapes of classes of various difficulty levels are available.
• The output is a two-dimensional dataset.
• It is easy to use the generator in a batch mode by calling the function createDataset() with different parameters.

Getting started

See the folder examples for some demonstrational examples or run the script examples.m that generates them.

Publications

• Nejc Ilc, "Clustering Based on Weighted Ensemble," PhD thesis, University of Ljubljana, 2016.
• Nejc Ilc, "Data generator with a control over linear separability," Proceedings of 26th International Electrotechnical and Computer Science Conference, Portorož, Slovenia, September 2017. Ljubljana: IEEE Region 8, the Slovenian section of IEEE.

Acknowledgements

In this project we reused the code from:

• Michael Chen (sqdistance2.m)
• Nicolo Giorgetti and Niels Klitgord (A Matlab MEX Interface for the GLPK library)