[1N5-GS-13] AI application: Machine learning and application (2)

[1N5-GS-13-02] Prioritizing Data Acquisition for Bayesian Networks with Variable-Specific Confidence Level Sensitivity Analysis

〇Megumi Hori¹, Aaron Bramson^1,2,3,4 (1. GA Technologies Inc., 2. RIKEN Center for Biosystems Dynamics Research, 3. Department of General Economics, Ghent University, 4. Department of Software and Information Systems, University of North Carolina Charlotte)

Keywords:Bayesian Network, Scoring, Confidence Levels, Sensitivity Analysis, Data Acquisition Strategy

Bayesian networks capture how the change in one variable affects others using conditional probabilities and repeatedly applying the belief propagation method to determine the probabilities of each state. Bayesian networks can be used as scoring engines that consider the relationships among explanatory variables. However, because data includes uncertainties (e.g. missing data, insufficient resolution, small sample size) it is necessary to judge the confidence of the score considering both the uncertainties and the relationships in the data. In this paper, probabilities and confidence levels are calculated in parallel at each node so that the sensitivity of the confidence levels reflects the probabilistic dependencies among nodes. Thus, input nodes with the greatest probabilistic impact also have a proportionally larger impact on the confidence of the score. In this way we can (1) identify the variables with the greatest effect on the confidence of the score, (2) increase confidence by reducing the uncertainty of important variables, and (3) enhance model robustness by prioritizing data. We demonstrate the proposed method using a simplified system that scores areas by overall safety.