Bayesian networks in survey data: Robustness and sensitivity issues

Bayesian networks (BN) implement a graphical model structure known as a directed acyclic graph (DAG) that is popular in statistics, machine learning, and artificial intelligence. They enable an effective representation and computation of a joint probability distribution (JPD) over a set of random variables. The paper focuses on the selection of a robust network structure according to different learning algorithms and the measure of arc strength using resampling techniques. Moreover, it shows how 'what-if' sensitivity scenarios are generated with BN using hard and soft evidence in the framework of predictive inference. Establishing a robust network structure and using it for decision support are two essential enablers for efficient and effective applications of BN to improvements of products and processes. A customer-satisfaction survey example is presented and R scripts are provided.