关键词: 群体性突发事件, 因果贝叶斯网络, 预警模型

Abstract: Mass emergencies can be regarded as complex disaster systems, characterized by numerous system elements and intricate direct and indirect relationships among these elements. Furthermore, the situation of mass emergencies often evolves rapidly, exhibits diverse forms, and possesses significant uncertainty in their development and impact outcomes. The aforementioned system complexity and uncertainty of mass emergencies pose challenges to event early warning, necessitating a method that can systematically represent all elements and their complex interrelationships while quantifying the uncertainty of these elements. To address this issue, this paper constructs an early warning model for mass emergencies based on causal Bayesian networks. This model can effectively utilize missing data, accurately identify and represent the complex causal relationships between elements, and quantify the uncertainty of variables through conditional probabilities. Consequently, by means of probabilistic inference, it achieves the prediction and early warning of outcomes and warning levels. Therefore, this model can simultaneously resolve the system complexity and uncertainty of mass emergencies, enhancing the accuracy, relevance, and interpretability of early warnings. Moreover, the model presented in this paper not only intuitively outputs the event consequences and warning levels, addressing what the question is, but also answers the key causes leading to these outputs, addressing why the question is. This thereby provides more effective decision support for the emergency management of mass emergencies.
To simplify the complexity of mass emergencies, this paper first introduces the disaster system theory of “disaster-prone environment-causative factors-vulnerable bodies” to construct a representation model for the disaster system of mass emergencies, providing the initial structure for constructing the causal Bayesian network. Subsequently, based on the representation model, the system elements of mass emergencies are identified to provide variables and data for the construction of the early warning model. Finally, the greedy fast causal inference (GFCI) algorithm is selected for structure learning to solve the pseudo-causality problem brought about by unobservable latent variables and confounding variables; the expectation-maximization (EM) algorithm is used for parameter learning to address the issue of missing data, ultimately constructing a causal Bayesian network model for the early warning of mass emergencies. Through probabilistic inference with the CBN, the possible outcomes and warning levels of mass emergencies are output, and based on sensitivity analysis, the most critical causal variables leading to the outcomes of mass emergencies are identified.
This paper collects 1138 complete cases of mass emergencies from the Wise News database to construct a causal Bayesian network for the early warning of mass emergencies and validate its effectiveness. The results show that the overall accuracy of the model reaches 0.92, and various indicators demonstrate that the model constructed in this paper achieves good results in predicting an early warning of mass emergencies. Additionally, the sensitivity analysis results show that the most critical variable affecting the intensity of online public opinion is “road traffic congestion,” the key variable affecting casualty numbers is “assault on police officers,” and the most critical variable affecting economic loss is whether “key departments” are damaged. Emergency management departments are recommended to focus on these variables when making decisions. Furthermore, this paper analyzes the evolutionary mechanisms of different types of mass emergencies and finds that events related to interest demands and land acquisition and demolition are more likely to result in “economic losses” and “casualties,” as they are directly related to people’s interests. In contrast, events related to social indignation and social disputes are more likely to lead to high-heated online public opinion consequences.
This study primarily relies on historical case data of mass emergencies. Due to the difficulty of obtaining time attributes, the dynamic characteristics of mass emergencies are not considered. In the future, we will further enrich the case data with time attributes and construct an early warning model for mass emergencies based on dynamic causal Bayesian networks to achieve more efficient and accurate results.

Key words: mass emergencies, causal Bayesian network, warning model