Dynamic Trend Detection in U.S. Border Security Social-Media Networks

Detecting temporal trends in large networks has strategic importance in many domains, such as cybersecurity and social media analytics, where the activities of key actors (e.g., activists, terrorists, leaders) are concerned. In a large evolving network, the relationships of actors in the network often change over time. Characterizing these changes can provide important insight on individualand group-level activities. This insight can inform situational understanding and intelligence analysis in the cyber domain. This research developed and validated a dynamic network activity model to characterize temporal trends in a large social-media network of interactive human agents. The model supports prediction of agent activities over time through modeling agents’ network interactions and network growth. We argue that large social-media networks exhibit significant effects of randomness and exponential growth due to community size, low connection cost, and high reachability. To study its predictive accuracy, the model was compared against an existing model that is based on exponential aggregation of agent activities. The two models were validated using a social-media community focused on U.S. border and immigration security. The community consists of 210,921 human agents who posted 533,246 messages and formed 453,552 links among agents. Temporal networks were extracted from the community, where each network captures a pre-defined temporal length of activities. Each model was used to predict activities of human agents given their historical activity levels. We implemented these prediction using Apache Spark, a distributed bigdata platform, and its graph computation package, GraphX. The experimental results show that the proposed model achieved significantly better accuracy than the baseline model. This research should contribute to providing new approaches and system artifacts for dynamic trend detection in social-media networks, reporting new findings of network trend detection, and providing new technical approaches to process large graph-based data.