Summary (Рандомизированные алгоритмы на основе интервальных узорных структур)

as a manuscriptMasyutin AlexeyINTERVAL PATTERN STRUCTURES RANDOMIZEDALGORITHMSFOR CLASSIFICATION AND REGRESSION TASKSIN CREDIT RISK MANAGEMENTPhD Dissertation Summaryfor the purpose of obtainingPhilosophy Doctor in Computer Science HSEMoscow –– 2018The PhD Dissertation was prepared at National Research University Higher School ofEconomics.Academic Supervisor::Sergei O. Kuznetsov, Doctor of Science, professor ofNational Research University Higher School ofEconomicsPhD Dissertation Relevance. The development of information technologiescreates a much tougher competitive environment for banks and credit institutions.For example, since January 2017, Russia's two largest telecommunicationscompanies have begun to provide loans to their customers, although they have neverdeveloped this business line before1.

The provision of such services on the part ofnon-financial companies was made possible precisely through the introduction of amodern IT infrastructure to store a large amount of data on customers and usemathematical modeling techniques to assess the creditworthiness of customers. Theinternational leader in the sphere of strategic consulting services, McKinsey,publishes studies according to which by 2025 the world banking industry willundergo significant changes. One of the main reasons for the transformation is thedigitalization of banking services, the rapidly growing volume of data on customersand their operations, the emergence of new types of risk associated with the use ofmachine learning in decision-making.Mathematical modeling in the banking sector finds one of its main applicationsin the field of risk management.

A prerequisite for effective risk management is theability to adequately assess the probability and magnitude of the risk. At the moment,risk assessment problems are approached by a wide range of statistical tools, such asscorecards, rating models, loss given default modeling. At the same time one of keyreasons for the numerous reviews of licenses is, among other things, the low qualityof the loan portfolio; inadequate estimates of the probability of default and / or lossgiven default 2 (for example, Probusinessbank, Tatfondbank in Russia). So, in thepress releases of the Central Bank of Russia, one can find following statements:"with unsatisfactory asset quality credit institution inadequately assessed the risks","as a result of expected loss calculation given the risks accepted, the credit institutioncompletely lost its capital", etc.With the increase in the volume of customer data, standard methods ofclassification and regression are yielding lower accuracy in comparison to morecomplex algorithms, such as random forests, support vector machines, and decisiontrees.

Nevertheless, risk management requires the property of interpretability of thepredictions received by the model, which in the case of complex algorithms isusually impossible. In addition, there are certain prescriptions of the Bank of Russiathat regulate the process of assessing credit risk on the basis of mathematical models,and banks undergo a detailed procedure for validating the use of models for riskassessment, a significant part of which is to test the stability of the model and itsbusiness logic.1http://www.rbc.ru/finances/11/01/2017/587500529a794767fa723fa8, abovementioned companies are Beeline and MTS.2http://www.cbr.ru/press/PR/?file=12082015_085127ik2015-08-12t08_46_23.htmThis PhD Dissertation solves the problem for interpreted decision rules thatoutperform classification and regression methods conventional in the banking sphere,such as scorecards and decision trees, but at the same time preserve theinterpretability of the forecasts produced.This is achieved by formal concept analysis (FCA) and interval patternstructures technique.

Several new definitions and modifications were introduced sothat one could be able to perform continuous target variable prediction via intervalpattern structures and to handle datasets with considerable number of observations.Object of Research is interpretable mathematical models for probability ofdefault and loss given default estimation.PhD Dissertation Goal is the development of innovative methods for creditscoring and loss given default estimation which provide higher accuracy incomparison to credit scorecards and decision trees, provided the methods areinterpretable from business logic perspective.The two main novelties of PhD Dissertation are as follows:1. Query-Based Classification Algorithm which provides randomizedprocedure to solve credit scoring problem for datasets with large numberof observations;2. Adopting tools of formal concept analysis to regression task (i.e.

case ofcontinuous target variable) by developing Query Based RegressionAlgorithm. The algorithm allows one to solve loss given default problem.PhD Dissertation offers data analysis algorithms that have accuracy superiorto simple algorithms widely adopted within the banks (such as logistic regression,decision trees and scorecards) and still maintain the property of interpretability insense that they provide a decision maker with a set of rules applicable to theborrower creditworthiness assessment. In order to achieve this goal several noveltieswithin the methods of formal concept analysis (FCA) and interval pattern structureswere introduced.

The reasons why FCA methods are suitable for credit riskassessment under the interpretability requirements are discussed in the main text ofthe dissertation. The novelty brought to the well-developed tools of FCA consists oftwo parts.The first one is that FCA is adopted to credit scoring problem based onnumerical data with the step of concept lattice construction being omitted (querybased classification or "lazy" classification).

This allows one to work with thedatasets with large number of observations which is vital for banks as soon ashistorical data is typically large.The second one is that we introduce a modification to FCA method based oninterval pattern structures which allows one to solve regression problem. To ourknowledge FCA methods were not applicable to such type of data analysis problembefore. The crucial difference in regression problem is that the target variable isdistributed continuously. The proposed method is designed to solve loss givendefault prediction problem.The Practical Value is justified by experiments on the comparativeevaluation of the different algorithms for classification and regression problems onreal bank data as well as open data.

All proposed methods are implemented assoftware packages designed to solve credit scoring problems.Proposed methods and algorithms were applied as a test project to thecompany's two client datasets of top-10 Russian bank and results of accuracycalculations and benchmark analysis are provided in the dissertation.The reliability of the obtained results is justified by application ofmathematical models and methods, by experiments comparing the results ofapplying the traditional methods developed.Publications.First-tier publications:1. Masyutin A., Kashnitsky Y. Query-Based Versus Tree-BasedClassification: Application to Banking Data // Lecture Notes in ComputerScience (including subseries Lecture Notes in Artificial Intelligence andLecture Notes in Bioinformatics), 2017, 10352 LNAI, pp. 664-673.Second-tier publications:2.

Masyutin A. Credit scoring based on social network data // BusinessInformatics, No. 3 (33), 2015, pp. 15–23.3. Masyutin A. Alternative Ways for Loss-Given-Default Estimation inRetail Banking //Analysis of Images, Social Networks and Texts, 2014,Volume 436 of the series Communications in Computer and InformationScience, pp. 152-162.4. Masyutin A., Kashnitsky Y., Sergei O. Kuznetsov. Lazy Classication withInterval Pattern Structures: Application to Credit Scoring, in: Proceedingsof the International Workshop "What can FCA do for ArtificialIntelligence?" (FCA4AI at IJCAI 2015), CEUR Workshop proceedings,Vol.1430, p. 43-54.5.

Masyutin A., Sergei O. Kuznetsov. Continuous Target Variable Predictionwith Augmented Interval Pattern Structures: A Lazy Algorithm //Proceedings of the Thirteenth International Conference on ConceptLattices and Their Applications, pp.273-284, 2016.Other publications:1.

Masyutin A., Borisyuk V. Fraud Detection Scorecard // Risk-managementin credit organization, No. 2/2015. (in Russian)Discussions at conferences and seminars.1. 3th International Conference - Analysis of Images, Social Networks andTexts, AIST'2014, Ekaterinburg, Russia.Subject: "Alternative Ways for Loss-Given-Default Estimation in RetailBanking";2. 24th International Conference on Artificial Intelligence, Workshop "Whatcan FCA do for Artificial Intelligence?", FCA4AI at IJCAI 2015, Buenos Aires,Argentina.Subject: "Lazy Classification with Interval Pattern Structures: Application toCredit Scoring;3.

Research Seminar of the Graduate School of Computer Science, CS HSE,November 12, 2015.Subject: "Integrated risk management and machine learning tasks",(https://aspirantura.hse.ru/cs/announcements/164947571.html);4. 13th International Conference on the Concept of Lattices and TheirApplications, CLA'2016. Moscow, RussiaSubject: "Continuous Target Variable Prediction with Augmented IntervalPattern Structures: A Lazy Algorithm";5.DatabaseTechnologies,2016,Moscow,Russia(https://www.osp.ru/iz/tbd_dbms)Subject: "Credit scoring based on the analysis of formal concepts: personalmodels";6. 23rd International Symposium on Methodologies for Intelligent Systems,ISMIS'2017, Warsaw, Poland.Subject: "Query-based versus tree-based classification: application tobanking data";7. Seminar of ISSA, CS HSE, September 28, 2017.Subject: "Classification on demand based on interval descriptions in creditrisk management tasks" (https://cs.hse.ru/ai/issa/announcements/209647835.html).PhD Dissertation Contents.

This dissertation consists of 4 sections,conclusion and bibliography of 89 resources.The first section (introduction) covers the dissertation relevance, the problemsand tasks of the research, the subject of the research. Also, objectives of the work aredetermined, the main results are stated, theoretical and practical significance of thework is discussed.The second section describes the history of mathematical modeling withinbanking industry.