Bayesian analysis of sentiment surveys

Bayesian belief network (BN) models allow users to perform sensitivity analyses on a dependent/target variable given a combination of input variables. Such analyses on BN models are possible due to their holistic nature, interconnecting variables in a model through a combination of deductive and abductive inferencing. A sensitivity analysis provides the basis for hypothesis testing to extract rules from a BN model for explaining and characterizing the target variable with respect to input variables. However, determining which combinations of input variables have the most impact on a given target variable is challenging due to the combinatorial complexity of a large BN model containing hundreds of variables. Here we present an approach based on a statistical significance test that recursively generates combinations of evidence on input variables and prunes paths from the search tree that are unlikely to produce any results. To address the combinatorial complexity issue, we exclude those combinations of input variables from consideration that are d-separated from the target variable. To demonstrate the utility and scalability of our approach, we first extract a BN model semi-automatically from a corruption survey dataset on Afghanistan and then analyze and cluster sentiments from various provinces of that country applying the proposed sensitivity analysis technique. We also provide details of how the extracted rules are represented in an executable format to support decision making.