The task of the TreeSEEKER system is to develop learning models in the form of quasi-optimal decision trees. This system has implemented the following algorithms:

  • CsC (based on Czerwinski’s coefficient),
  • ID3/C4.5 (our implementation of classic Quinlan’a algorithm),
  • TVR (Tree-via-Rule), an algorithm that creates the decision tree using partial sequences of paths from the apparent root of the tree to selected leaf, and
  • a new algorithm called a VDP, searching for the most significant attribute in the set of attributes by classification of a data set.

       The VDP algorithm is based on generating belief networks with changed Dirichlet’s parameter. Let us note that the descriptive attribute which has the most important marginal influence on decision attribute is located in the root of tree. Developed decision trees are presented in to formats: textual and/or graphical forms. The quality of the developed tree is evaluated by means of two parameters. The first one (standard) is based on estimation of the error rate while unseen cases are classificated. The second one, the mean number of questions, is applied for the evaluation of the compactness of the tree. Important feature of the system is the ability to develop certain and possible learning  model for the analyzed datasets, provided that are some inconsistent data. 

An example of graphical representation of the decision tree