A Temporal Attention-based Model for Social Event Prediction

Large-scale social events like civil unrest, distinctly impacts our daily life. For this reason, it is essential to predict specific events in advance based on relevant information. Studies on data-driven event prediction assume that there are precursors for predictable events that can be tracked in history. Therefore, modeling prior evolution of a target event appropriately using relevant information or implicit indicators is of great importance to anticipate whether concerned events are likely to occur sometime in the future. However, there are issues among existing relevant studies: (I) how to properly define event data for training to match the realistic prediction scenario. (II) how to extract useful previous information from available data flow and model relevant evolution or dynamic feature of events reasonably. (III) it is both practical and urgent to mine precursors or clues from spatial-temporal data for interpreting prediction results. In this paper, we propose a novel feature learning framework for event prediction that can discover potential precursors from the input data. The prediction model primarily consists of a graph encoder module using GNN (Graph Neural Network) techniques and a temporal feature learning module employing attention mechanism. Meanwhile, we develop a backward tracking method to model the previous evolution of an event by retrieving prospective relevant events in the past. Multiple experiments conducted on datasets collected from various regions in the real world demonstrate appreciable performance of our proposed model in social event prediction task.