TY - JOUR

T1 - Bayesian belief network structure synthesis for risky behavior rate estimation

AU - Suvorova, A. V.

AU - Tulupyev, A. L.

PY - 2018/1/1

Y1 - 2018/1/1

N2 - Introduction: Studies in sociology, psychology, epidemiology, marketing or information security often face the issue of estimating the behavior rate (either individually or at the level of a population). Direct methods of behavior rate estimation are sometimes not available; hence, it is important to develop indirect methods. Earlier studies proposed an approach to risky behavior modeling based on a Bayesian belief network using the data about several last behavior episodes as its initial data. However, to apply this model in practice, we need to reduce its dependency from the initial experts’ assumptions about the relations between the elements of the model. Purpose: To propose a model modification which would not require an expert-based model structure, and to compare the modified model with the initial one. Methods: To test the model, we used an automatically generated dataset which followed some initial assumptions about the data. To form the structure of a Bayesian belief network, we used a score-based hill-climbing algorithm with Bayesian information criterion score. Results: We proposed to modify the approach to risky behavior modeling in terms of Bayesian belief network based on the data about several last behavior episodes. The initial expert-based model and the model with a data-based structure were compared. Formal scores were better for the data-based structure, while the prediction quality was slightly better for the expert-based model. Hence, we can use both these models for practical applications; the choice depends on the assumptions and limitations of a particular task.

AB - Introduction: Studies in sociology, psychology, epidemiology, marketing or information security often face the issue of estimating the behavior rate (either individually or at the level of a population). Direct methods of behavior rate estimation are sometimes not available; hence, it is important to develop indirect methods. Earlier studies proposed an approach to risky behavior modeling based on a Bayesian belief network using the data about several last behavior episodes as its initial data. However, to apply this model in practice, we need to reduce its dependency from the initial experts’ assumptions about the relations between the elements of the model. Purpose: To propose a model modification which would not require an expert-based model structure, and to compare the modified model with the initial one. Methods: To test the model, we used an automatically generated dataset which followed some initial assumptions about the data. To form the structure of a Bayesian belief network, we used a score-based hill-climbing algorithm with Bayesian information criterion score. Results: We proposed to modify the approach to risky behavior modeling in terms of Bayesian belief network based on the data about several last behavior episodes. The initial expert-based model and the model with a data-based structure were compared. Formal scores were better for the data-based structure, while the prediction quality was slightly better for the expert-based model. Hence, we can use both these models for practical applications; the choice depends on the assumptions and limitations of a particular task.

KW - Bayesian Belief Network

KW - Behavior Modelling

KW - Machine Learning

KW - Risky Behavior

KW - Structure Synthesis

UR - http://www.scopus.com/inward/record.url?scp=85058527879&partnerID=8YFLogxK

U2 - 10.15217/issn1684-8853.2018.1.116

DO - 10.15217/issn1684-8853.2018.1.116

M3 - Article

AN - SCOPUS:85058527879

VL - 2018

SP - 116

EP - 122

JO - ИНФОРМАЦИОННО-УПРАВЛЯЮЩИЕ СИСТЕМЫ

JF - ИНФОРМАЦИОННО-УПРАВЛЯЮЩИЕ СИСТЕМЫ

SN - 1684-8853

IS - 1

ER -