A Bayes-true data generator for evaluation of supervised and unsupervised learning methods

摘要: Benchmarking pattern recognition, machine learning and data mining methods commonly relies on real-world sets. However, there are some disadvantages in using data. On one hand collecting can become difficult or impossible for various reasons, the other variables hard to control, even problem domain; feature domain, where most statistical operate, exercising control is more hence rarely attempted. This at odds with scientific experimentation guidelines mandating use of as directly controllable observable possible. Because this, synthetic possesses certain advantages over In this paper we propose a method that produces guaranteed global class-specific properties. based overlapping class densities placed corners regular k-simplex. generator be used algorithm testing fair performance evaluation methods. strong properties researchers reproduce each others experiments by knowing parameters used, instead transmitting large