In the CrimeGraph project, I explored training a graph deep learning model to forecast the probability of crime events in the City of Baltimore, using data provided by SpotCrime.
The model's design was based on the Gated Localised Diffusion (GLDnet) architecture described in the 2020 paper Graph Deep Learning Model for Network-based Predictive Hotspot Mapping of Sparse Spatio-Temporal Events.
GLDnet models can learn from the spatial patterns of crime events by representing intersecting city streets as a graph of nodes and edges.
- Street Graph Feature Engineering
- Event Timestep Feature Engineering
- Building the Graph Localized Diffusion Network
- Model Training and Turning
The source code is available on GitHub.