Forecasting the daily outbreak of topic-level political risk from social media using hidden Markov model-based techniques

Nowadays, as an arena of politics, social media ignites political protests, so analyzing topics discussed negatively in the social media has increased in importance for detecting a nation's political risk. In this context, this paper designs and examines an automatic approach to forecast the daily outbreak of political risk from social media at a topic level. It evaluates the forecasting performances of topic features, investigated among the previous works that analyze social media data for politics, hidden Markov model (HMM)-based techniques, widely used for the anomaly detection with time-series data, and detection models, into which the topic features and the detection techniques are combined. When applied to South Korea's Web forum, Daum Agora, statistical comparisons with the constraints of false positive rate of <0.1 and timeliness of <0 show that, for accuracy, social network-based feature and, for sensitivity, energy-based feature give the best results but there is no single best detection technique for accuracy and sensitivity. Besides, they demonstrate that the detection model using Markov switching model with jumps (MSJ) with social-network based feature is the best combination for accuracy; there is no single best detection model for sensitivity. This paper helps make a move to prevent the national political risk, and eventually the predictive governance benefits the people.