A Method Based on Entity Interaction Graph for Detecting Social Events Using GDELT

Social events are group activities gathered at a certain time and location. Timely detecting social events is of great significance to grasp the dynamics of social development. Most of the existing researches on social event detection are oriented to unstructured data, and these methods are difficult to applied for structured data. Considering that structured data focuses on the event level and has strong authority, it is necessary to put efforts into related research on structured data. However, the existing researches on structured data rely heavily on rule-based feature construction, which cannot fully mine the hidden layer features representing events. Fortunately, deep learning methods can solve this problem. In this paper, we design a social event detection method based on entity interaction graph mining to achieve this task. Firstly, we build a feature mining module based on multi-relational graph convolutional network, and update the feature information by the powerful feature mining ability of graph convolutional network. Then, a feature selection module is designed based on the graph pooling mechanism to select important nodes and edges. Finally, the feature concatenation module is constructed, the feature mining layer and feature selection layer are reasonably laid out, the embeddings of nodes and edges of corresponding layers are obtained by max pool and average pool, and the embeddings of corresponding layers are concatenated before classification. Finally, a classifier is constructed to obtain the final output. In order to demonstrate effectiveness and performance of our model, we construct datasets and design evaluation experiments on this basis. Experimental results show that our model outperforms the baselines with better generalization.