Models for respondents' behavior rate estimate: Bayesian Network structure synthesis

Abstract

The paper described the structure comparison of Bayesian Belief Network models for individual behavior rate estimate based on data about the last episodes of that behavior. We compared two types of network structures: expert-based and data-based. For model learning and evaluation we used data from social network VKontakte about episodes of publishing posts. The sample size was 3803 users with 785066 posts in total for the half-year period; training dataset included 75% of the sample, the rest 25% were used as test dataset. The data-based structure represented slightly better quality scores, while prediction quality was almost the same: 90.5% accuracy for model with expert-based structure and 89.6% accuracy for data-based model. Hence, both models showed quite similar results that allows, for example, reducing computations and applying expert-based structure for solving practical issues.