Predictive Graph Convolutional Networks - 19-008

Metron and Northeastern University propose to design, develop, and validate a proof-of-concept predictive Graph Convolutional Network (GCN) capability using open source Reddit and GDELT data. We propose: (1) to extract and preprocess open-source Reddit and GDELT data, (2) to design a predictive graph convolutional neural network model, (3) to implement and train that model, and (4) to validate the predictive capability of the model. Northeastern University brings experience developing the state-of-the-art Diffusion Convolutional Recurrent Neural Network (DCRNN) model, a GCN-RNN (Recurrent Neural Network) hybrid for forecasting traffic from a sequence of directed graphs representing historical traffic flows. Metron brings extensive experience in developing, implementing, and validating machine learning models, including GCN and RNN models designed by the Principal Investigator, and transitioning them by integrating them into existing government systems with accompanying interactive user interfaces. In addition to DCRNN, Northeastern University also brings expertise on other GCN and RNN models.