Conference Program Home
  My Program

All Times EDT

Abstract Details

Activity Number: 245 - Bayesian Models for Clustering and Latent Allocation
Type: Contributed
Date/Time: Tuesday, August 9, 2022 : 8:30 AM to 10:20 AM
Sponsor: Section on Bayesian Statistical Science
Abstract #323394
Title: Graph Product Partition Models
Author(s): Changwoo J Lee* and Huiyan Sang
Companies: Texas A&M University and Texas A&M University
Keywords: Graph partition; Community detection; Generalized Bayesian inference; Random spanning tree

We propose the Graph Product Partition Model (graphPPM), a probability model for random partitions of nodes set on a fixed graph where each subset of nodes induces a connected subgraph. The graphPPM has a product form of cohesion functions, which encodes information on the internal and external connectivity of each subgraph induced by a graph partition. For the community detection task with node attributes, graphPPM can be used as a prior distribution on graph partition which directly incorporates graph structures and hence bypass the need to impose graph generative assumptions such as the stochastic block model. We propose two novel inference strategies: 1) the importance sampling method for estimating the pairwise similarity matrix of the graph partition, and 2) the split-merge sampler using uniform random spanning trees to design the MCMC transition kernel on the partition space. We demonstrate the effectiveness of graphPPM in real networks for community detection tasks.

Authors who are presenting talks have a * after their name.

Back to the full JSM 2022 program