Skip to content

Latest commit

 

History

History
20 lines (14 loc) · 1.86 KB

README.md

File metadata and controls

20 lines (14 loc) · 1.86 KB
Raw

MINOTAUR

Multi-objectIve evolutioNary algOrithm for consisTent and interpretAble mUlti-label Rules is a .NET Core implementation of genetic algorithm that generates a collection of rule-based classification models, each offering a different compromise between predictive power and interpretability.

The default fittest selection algorithm is NSGA-II. The default metrics being optimized are F-Score (in the training dataset), number of rules in the model and the average "feature-space volume" covered by the model's rules.

Command line arguments

  • This readme file may be out of date; to see the current command line arguments, check this file.

Input files format

  • The input files must be in the 'csv' format.
  • Columns must be separated by a ","
  • Rows must be separated by a 'endline' character (depends on the operating system)
  • All values must be finite floating point values (read the Current Limitations section)

Current limitations

  • The current implementation forces all rules to have a feature test for each feature, even if the test is useless, e.g.: -infinity <= feature[x] < +infinity. To accurately calculate interpretability-related metrics for final models generated by the algorithm, such cases should be post-processed and removed.
  • The current implementation only support continuous features; if the dataset contains categorical features they won't be parsed correctly. In case they are successfully parsed, they will be treated as continuous features, which means the results will probably be nonsensical.
  • The current implementation only supports the maximization of the fitness values; therefore objectives that must be minimized (e.g. model size) are multiplied by -1.