The neural network classification problem approach based on the relation-features

The research is devoted to developing classification problem approaches for complicated objects. The target of the research is increasing the quality of classification systems (pattern recognition systems). By complicated object we mean object that have a few features yielded little information and can be considered as complex system with its features as subsystems. A concise classification problem statement for complicated objects based on the relation-features is defined. The approaches are based on the relations (with a different number of \"places\") on the sets of the probabilistic features of instances. We introduced the concept of "relation-features": thecartesian products of dyads or triplets of the probabilistic features of instances (this is dyadic or binary relation and triadic or ternary relation respectively). We added steps of features preconditioning during the stages of training and classifying (solving the clustering subproblem) to form these "relation-features". Then we can use them in any classification algorithms. The algorithms of using the approach and decision support tools to solve classification problems in medicine are developed. Neural network procedures and statistical classification methods are applied and compared with breast cancer diagnostics using features we got from some medical tests. Experimental results show that both the neural network and statistical classification problem approaches based on the "relation-features" allows the quality of classification problem solution (precision and recall) increase.