BSTPP: a python package for Bayesian spatiotemporal point processes

Isaac Manring, George Mohler, Xenia Miscouridou

Abstract: Spatiotemporal point process models have a rich history of effectively modeling event data in space and time. However, they are sometimes neglected due to the difficulty of implementing them. There is a lack of packages with the ability to perform inference for these models, particularly in python. Thus we present BSTPP a python package for Bayesian inference on spatiotemporal point processes. It offers three different kinds of models: space-time separable Log Gaussian Cox, Hawkes, and Cox Hawkes. Users may employ the predefined trigger parameterizations for the Hawkes models, or they may implement their own trigger functions with the extendable Trigger module. For the Cox models, posterior inference on the Gaussian processes is sped up with a pre-trained Variational Auto Encoder (VAE). The package includes a new flexible pre-trained VAE. We validate the model through simulation studies and then explore it by applying it to shooting data in Chicago.