Violence Diffusion Prediction via Deep Learning

Kunpeng Zhang, Assistant Professor
Robert H. Smith School of Business

This project seeks to reduce or prevent violent crimes by using large-scale data collection from Twitter and a methodology called Recurrent Cascades Convolutional Networks (CasCN) to monitor, predict, and visualize violent cascades in a given area. Using user-generated data from Twitter, this research could help proactively offer information for law enforcement policies regarding violence.

The research implements real-time social media monitoring to dynamically predict the cascade size of a certain violent topic at a specific geographical area (e.g., if a robbery happened in College Park today, how many residents in this area will pay attention to it for the next three days?), in addition to standard reporting. The research leverages long short-term memory (LSTM) and graph convolutional network (GCN) to predict the future size of a given cascade.

Deep learning is a machine learning technique that teaches computers to do what comes naturally to humans: learn by example. Deep learning is used in technologies such as autonomous cars and voice-controlled devices. Using a deep-learning framework, the researcher is modeling the structural and temporal characteristics of violence, as well as the features. The algorithm is being implemented and evaluated using Python and Pytorch.

Next steps include using existing data to test and val