DGCN-rs: A Dilated Graph Convolutional Networks Jointly Modelling Relation and Semantic for Multi-event Forecasting

Forecasting multiple co-occurring events of different types (a.k.a. multi-event) from open-source social media is extremely beneficial for decision makers seeking to avoid, control related social unrest and risks. Most existing work either fails to jointly model the entity-relation and semantic dependence among multiple events, or has limited long-term or inconsecutive forecasting performances. In order to address the above limitations, we design a Dilated Graph Convolutional Networks (DGCN-rs) jointly modelling relation and semantic information for multi-event forecasting. We construct a temporal event graph (TEG) for entity-relation dependence and a semantic context graph (SCG) for semantic dependence to capture useful historical clues. To obtain better graph embedding, we utilize GCN to aggregate the neighborhoods of TEG and SCG. Considering the long-term and inconsecutive dependence of social events over time, we apply dilated casual convolutional network to automatically capture such temporal dependence by stacked the layers with increasing dilated factors. We compare the proposed model DGCN-rs with state-of-the-art methods on five-country datasets. The results exhibit better performance than other models.