Explore a dynamic feature allocation model for time-stamped document collections in this 30-minute talk by Valerio Perrone from the Oxford-Warwick Statistics Programme. Delve into the Wright-Fisher Indian Buffet Process (WF-IBP), a Bayesian nonparametric model that extends the static Indian Buffet Process to account for changing topic popularity over time. Learn how this model is applied to develop a nonparametric focused topic model for analyzing the full corpus of NIPS papers from 1987 to 2015. Gain insights into the motivation, mathematical foundations, and practical applications of the WF-IBP, including its development based on population genetic models and Poisson random fields. Compare dynamic and static models using simulated data and real-world NIPS data, and understand the advantages of this approach in capturing evolving topics in large-scale document collections.
Overview
Syllabus
Introduction
Motivating Example
Poisson Random Field Development based on a population genetic model Sawyer and Hartl, 1992
Background: Indian Buffet Process
Background: Beta Process
The Wright-Fisher Model
The Wright-Fisher Diffusion
The Poisson Random Field
Poisson Random Field for Indian Buffet Processes
The WF-IBP model
MCMC inference
Simulated Data with Linear-Gaussian Observation Model
WF-IBP Topic Model
Comparing Dynamic vs Static Models (Simulated Data)
Comparing Dynamic vs Static Models (NIPS Data) Test-set perplexity
NIPS Topic Model
Concluding Remarks
References
Taught by
Alan Turing Institute
