Measures for the characterisation of pattern-recognition data sets
Christiaan M Van der Walt, Etienne Barnard · 2007
We study the relationship between the properties of data and classifier performance. Data measures are employed to characterise classification problems and it is shown that these data measures successfully capture important characteristics of the relationship between data and classifiers. The proposed data measures can be used to predict the classification performance of real-world data sets and to gain insight into the structures and properties of real-world data. 1.