Reproduce “Spatiotemporal Clustering with Neyman-Scott Processes via Connections to Bayesian Nonparametric Mixture Models” on OpenResearch

About this paper

Yixin Wang, Anthony Degleris, Alex H. Williams, Scott W. Linderman

This work establishes a formal connection between Neyman-Scott processes (NSPs) and Bayesian nonparametric mixture models, developing a collapsed Gibbs sampling algorithm with split-merge moves. The approach provides a scalable and efficient Bayesian inference method for spatiotemporal clustered data, outperforming traditional RJMCMC in convergence and demonstrating effective application in neural sequence discovery and document event detection.

Key insights

• Neyman-Scott processes with gamma-distributed weights can be formally viewed as a special case of Mixture of Finite Mixture Models.
• NSPs can bridge between Mixture of Finite Mixture Models and Dirichlet Process Mixture Models, approaching the latter in a specific limit where the number of latent events grows.
• The generative process of NSPs induces an exchangeable random partition and a Pólya urn scheme, enabling efficient collapsed Gibbs sampling by allowing marginalization over latent event properties.

Summary by alphaXiv.