THE ALGORITHM FOR BINARY CLASSIFICATION ON GRAPH-BASED DECISION-MAKING IN THE TASKS OF CREDIT SCORING 

Alexey N. Kislyakov, Candidate of technical sciences, associate professor of sub-department of information technology, Vladimir branch of the Russian Academy of National Economy and Public Administration under the President of the Russian Federation (59a, Gorky street, Vladimir, Russia), E-mail: ankislyakov@mail.ru 

519.17 

10.21685/2227-8486-2021-1-3 

Background. The work is devoted to the actual problem of constructing decision graphs of an optimal structure, which are used to solve problems of binary classification and create predictive models of socio-economic indicators. The aim of the work is to generalize the experience of constructing decision trees and graphs and to study the quality of classification models based on them.
Materials and methods. Examples of implementation of algorithms based on trees and randomized ensembles of oriented acyclic decision graphs (DAG or jungle of decisions) for the problem of credit scoring as one of the directions of modification of ensemble algorithms based on decision trees are shown. The main difference between decision graphs and decision trees is the presence of tree nodes that are binary classifiers and can be connected to other nodes that are not hierarchically connected to the parent node. Thus, a binary graph compared to a decision tree can have not only root nodes, but also split nodes, as well as leaves.
Results. Since the features corresponding to different classes have different values of the indicator, the root node can divide them according to this feature, as well as the child nodes can do it. Based on the entropy index of the tree, an information gain can be calculated, which will allow optimizing the graph structure by minimizing the total weighted entropy of a set of predictor values. As a result, not only the separation rules are developed, but also the rules for connecting features within the tree, so with a lower depth of construction of the decision graph, it has a greater ability to describe the relationships of the feature space of a complex system.
Conclusions. The possibility of overcoming the problem of instability of finite predictions of models based on decision trees on relatively small data samples by switching to graph classification models are describe. Studies have shown that decision graphs are the most efficient classification algorithm that shows the best results on a small training sample. 

decision trees, decision graphs, binary classification, machine learning 

Kislyakov A.N. The algorithm for binary classification on graph-based deci- sion-making in the tasks of credit scoring. Modeli, sistemy, seti v ekonomike, tekhnike, prirode i obshchestve = Models, systems, networks in economics, technology, nature and society . 2021;1:29–41. (In Russ.). doi:10.21685/2227-8486-2021-1-3