The Idea
The concept of the project was inspired by the problem of contaminated datasets. Such datasets include not only proper patterns that we wish to process (classify), but also garbage patterns that need to be rejected prior to any processing.

Practical Applications
The problem of contaminated datasets arises commonly when patterns are acquired automatically and are preprocessed before data mining is carried out.

Our Focus
Our focus is on pattern recognition with garbage patterns rejection applied to:
  • optical character recognition (supervised learning),
  • music notation recognition (supervised learning for imbalanced dataset),
  • computer-aided medical diagnosis (unsupervised learning).
Project Name
Classification with Rejection

The National Science Center, grant No 2012/07/B/ST6/01501, decision no DEC-2012/07/B/ST6/01501

Involved Institutions
Systems Research Institute, Polish Academy of Sciences, Warsaw, Poland
Warsaw University of Technology, Warsaw, Poland
University of Alberta, Edmonton, Canada

The most important and representative objectives of the project involve:
  • to formulate the complete concept of pattern recognition with rejection,
  • to construct training and test sets for tasks of pattern recognition with rejection,
  • to develop classifiers with rejection based on standard machine learning methods, e.g. minimum distance, clustering, statistical methods and geometrical methods,
  • to develop methods for foreign patterns rejection based on imprecise information representation models,
  • to formulate and develop comprehensive methods for empirical assessment of classifiers.
Methods (highlights)
  • specifically trained collections of binary classifiers,
  • decision rules for unlabelled sets,
  • geometrical methods for feature space discrimination.