Computer program system BeliefSEEKER, applying some elements of rough sets theory, is capable of generating certain belief networks (for consistent data) and possible networks (for inconsistent data). Both types of networks can be converted onto respective sets of production rules called by us belief rules. The first step in the belief networks development is to load a decision table into the system. During the loading process, the system executes very detailed searching for erroneous and missing values. The program also informs user about number of contradictory, redundant and correct cases in the database. Additional element of processing is discretization of numerical attributes (it is worth notice that the BeliefSEEKER is capable to read discretized attribute ranges). The main tasks of the system are:

  • generating a network model,
  • probabilistic conclusion (i.e. generating probabilistic dependences between network nodes)
  • learning (i.e. a conditional probabilities estimating by maximization of marginal probability)

       In the network model generating process different algorithms are used: K2, K2 (all combinations of descriptive attributes), Naive Bayesian Classifier, and reversed Naive Bayesian Classifier. The structure of generated model is presented in graphical form. System provides complete interaction with the user. The one belief network with arbitrarily selected Dirichlet parameter can be generated, or – according to the user’s needs – a set of belief networks are developed in the incremental change of this parameter. Some of the networks are retained and applied in the classification process of unseen cases. Correctness of the network is checked in the classification module. This module identifies incorrect classifications in the matrix of distraction.