Reproduce “Neural Point Process for Learning Spatiotemporal Event Dynamics” on OpenResearch

About this paper

Zihao Zhou, Xingyi Yang, Ryan Rossi, Handong Zhao, Rose Yu

DeepSTPP presents a neural point process model that learns continuous spatiotemporal event dynamics by leveraging a neural latent process and non-parametric kernels. This design allows for a closed-form likelihood integration, enabling it to achieve superior prediction accuracy for event times and locations while training 100 times faster than state-of-the-art baselines.

Key insights

• Integrating a neural latent process with non-parametric kernel functions can effectively model complex, irregular spatiotemporal event dynamics while maintaining analytical tractability for likelihood calculations.
• Jointly learning continuous spatial and temporal event dependencies with deep neural networks significantly improves predictive capabilities over models that treat these aspects separately or discretely.
• Designing the intensity function for closed-form integration dramatically boosts computational efficiency, enabling large-scale application of sophisticated spatiotemporal point process models.

Summary by alphaXiv.