Bayesian network (BN)has been widely used in modeling expert knowledge and reasoning in many application domains, due to its power of representing probabilistic knowledge over a set of interacting variables under significant uncertain. However, the actual construction of a Bayesian network model for a domain remains to be a challenging task since it involves learning the causal structure and the conditional probability table of BN variables from experts, and/or observational data. In this paper, we explore the unique challenge of applying Bayesian network to model how public safety officials represent and reason about threats of evolving mass protest. Anticipating threats of mass protests is crucial for choosing proper intervention strategies, but it is also inherently difficult because a large number of interacting factors contribute to the dynamics of protests with high uncertainty and contingency. In constructing the BN model for this complex domain, we found that traditional methods of discovering BN structure either from data or experts are inadequate, due to the scarcity of data and the highly stochastic nature of mass protest events. Instead, we proposed a hybrid approach (called “ISM-K2”)which enhances the BN structure learning methods (K2 algorithm)by expert knowledge elicited using the ISM (interpretive structural model)method. We show that the BN model constructed using ISM-K2 approach is superior than three other base-line models (the logistic regression model, the BN constructed only by expert knowledge from ISM, and the BN constructed only by data learning). Finally, we show the potential of using our BN threat assessment model for supporting practical policing decisions.
All Science Journal Classification (ASJC) codes
- Computer Science Applications
- Artificial Intelligence