K-plus-nearest neighbor method development for credit scoring machine learning tasks

Abstract

The pace of development of modern risk management and data mining technologies causes the relevance of searching for new or improved effective methods for statistical and non-statistical forecasting, as well as forming the problems of deep study of existing methods and characteristics of their application conditions. Machine learning, namely memory-based learning is one of the most practically useful, broad and insufficiently studied areas. Also, the development of modern information technologies and ways to improve readability and simplicity of code causes the relevance of the study support with the implementation of the fourth-generation programming language. The research deals with developing basic and advanced k-plus-nearest neighbor method as significantly improved classical k-nearest neighbor method with eliminated shortcomings and inaccuracies of practical realization: the problem of selecting a metric space and the metrics itself, problem of using categorical (including sampled) variables on the set, the issue of probabilistic classification, problem of taking into account equally spaced groups of elements relative to the element to be classified, the model optimality criterion based on the method and the method of its use for selecting the optimal parameter, ways to accelerate application. The main work is focused on using the methodology and indicators of credit scoring in machine learning problems. The full code for the basic proposed method in the SQL language-MS SQL (T-SQL) dialect was given. As a result of the study, efficiency was determined at the stage of applying the basic proposed method in terms of the optimality criterion-Gini index relative to probabilistic forecasts compared to logistic regression in terms of two factors: the quality of forecasts and number of parameters to be optimized. The practical value of the results obtained on the example of simulation using mass consumer credit data lies in the simplicity and effectiveness of the proposed method by means only of the server part of the DBMS.