Bayesian Methods for Network & Image Analysis

Shuangjie Zhang Chair
UC Santa Cruz
 
Monday, Aug 5: 8:30 AM - 10:20 AM
5038 
Contributed Papers 
Oregon Convention Center 
Room: CC-G132 

Main Sponsor

Section on Bayesian Statistical Science

Presentations

Bayesian k-Space Estimation Decreases Image Noise and Increased Activation Detection

In fMRI, as voxel sizes decrease, there is less material in them to produce a signal, leading to a decrease in the signal-to-noise ratio and contrast-to-noise ratio in each voxel. There have been many attempts to decrease the noise in an image in order to increase activation, but most lead to blurrier images. An alternative is to develop methods in spatial frequency space. Reducing noise in spatial frequency space has unique benefits. A Bayesian approach is proposed that quantifies available a priori information about spatial frequency coefficients, incorporates it with observed spatial frequency coefficients, and estimates spatial frequency coefficients values a posteriori. Inverse Fourier transform reconstructed images form marginal posterior mean estimated spatial frequency coefficients have reduced noise and increased detection power. 

Keywords

Bayesian

k-space

fmri

imaging 

View Abstract 2490

First Author

Dan Rowe, Marquette University

Presenting Author

Dan Rowe, Marquette University

Hierarchical DP ERGM Mixtures of Heterogeneous Network Data

A concern when modeling ensembles of networks obtained from diverse sources is the presence of heterogeneity in the determinants of network structure; a property captured well by Dirichlet Process mixtures. Recently, work using DP mixtures of exponential family random graph models has shown that the approximate likelihood can be successfully employed for posterior inference on hidden populations, enabling broader families of such mixtures. Here, we consider hierarchical DP ERGM mixtures with both partial pooling (effects treated homogeneously across networks) and heterogeneity in included effects (allowing different effects between different mixture components). The former allows for semi-parametric estimation of given effects of interest (controlling for heterogeneity in other effects), while the latter allows for model averaging (can be used for model selection & clustering applications). The prior on the concentration parameter allows us to assign minimal prior weight to undesirable size distributions while adaptively assessing hidden population sizes. We evaluate the behavior of extended DP ERGM mixtures via simulation and display an application to heterogeneous network data. 

Keywords

Networks

Exponential Family Random Graph Models

Dirichlet Process

Bayesian statistics

Hierarchical Modelling

Data Clustering 

View Abstract 3144

Co-Author

Carter Butts, University of California-Irvine

First Author

Frances Beresford, University of California Irvine

Presenting Author

Frances Beresford, University of California Irvine

Bayesian Multi-scale Modeling and Monitoring of Large-Scale Agent Flows in Dynamic Networks

We discuss dynamic modeling of temporal paths of multiple agents through geographic networks with interest in detecting anomalous behavior of agents or subgroups of agents.The focus is sequential, monitoring incoming (noisy) time-trajectory data on huge numbers of agents on large networks in real time.Models represent geographical regions as networks of nodes as fine geographic zones.Interpretable Bayesian dynamic models are multi-scale in space, time and agents (ranging from individuals to subsets of agents defined via multiple intersecting features). Multi-scale models capture transitions over time of individual agents informed by activities at group levels. Methodology leverages decouple/recouples concepts, enabling model scalability and efficient computation.Analysis is overlaid with decision-theoretic monitoring customized to anomaly detection.Examples explore large-scale data arising from collaborative studies on real-time monitoring of human population movements in defined regions of a US city landscape.This highlights the roles and utilities of the methodology in characterizing normal dynamic patterns of trajectories at individual and group-levels, and in anomaly detection. 

Keywords

anomaly detection

Bayesian dynamic models

Bayesian model monitoring

multi-scale models

network flow modeling 

View Abstract 3117

Co-Author

Mike West, Duke University

First Author

Houjie Wang

Presenting Author

Houjie Wang

Comparing dependent Gaussian directed networks

We propose an approach to assess stability of Gaussian directed networks with networks constructed based on data from two time points. The proposed approaches unify network construction and comparison. A penalty introduced to the calculation of conditional posterior probability mass function for network differentiation ensures convergence to the underlying truth in probability. Simulations and real data applications support the feasibility of the method and the advantage of addressing dependence in network comparisons. 

Keywords

Gaussian network comparisons

Bayesian methods

Penalized conditional posterior probability

Variable selection. 

View Abstract 3324

Co-Author(s)

Xianzheng Huang, University of South Carolina
Hasan Arshad, University of Southampton

First Author

Hongmei Zhang, University of Memphis

Presenting Author

Hongmei Zhang, University of Memphis

Covariate-Assisted Bayesian Graph Learning for Heterogeneous Data

In a traditional Gaussian graphical model, data homogeneity is routinely assumed with no extra variables affecting the conditional independence. In modern genomic datasets, there is an abundance of auxiliary information, which often gets under-utilized in determining the joint dependency structure. In this talk, we consider a Bayesian approach to model undirected graphs underlying heterogeneous multivariate observations with additional assistance from covariates. Building on product partition models, we propose a novel covariate-dependent Gaussian graphical model that allows graphs to vary with covariates so that observations whose covariates are similar share a similar undirected graph. To efficiently embed Gaussian graphical models into our proposed framework, we explore both Gaussian likelihood and pseudo-likelihood functions. Moreover, the proposed model has large prior support. We show that based on the theory of fractional likelihood, the rate of posterior contraction is minimax optimal. The efficacy of the approach is demonstrated via simulation studies and an analysis of a protein network for a breast cancer dataset assisted by mRNA gene expression as covariates. 

Keywords

product partition model

Gaussian graphical model

pseudo-likelihood

G-Wishart prior

posterior contraction rate 

View Abstract 1889

Co-Author(s)

Yang Ni, Texas A&M University
Debdeep Pati, Texas A&M University
Bani Mallick, Texas A&M University

First Author

Yabo Niu, University of Houston

Presenting Author

Yabo Niu, University of Houston

Point Estimation of Networks from Posterior Samples

Bayesian networks are a method of modeling conditional dependencies among variables and have a wide variety of applications. Bayesian network models place a prior distribution on the network structure, and Markov chain Monte Carlo is typically used for model fitting, which results in thousands of networks sampled from the posterior distribution. Based on these samples, we propose a method to provide a point estimate of a Bayesian network structure. First, we introduce generalized structural Hamming (GSH) loss, a function between the adjacency matrices of networks which satisfies quasi-metric properties. We also introduce a stochastic sweetening algorithm to obtain a Bayes estimate by minimizing the Monte Carlo estimate of the posterior expected GSH loss using the available samples. We provide an investigation of existing methods and our proposed methods. Our loss function and search algorithm are implemented in an R package. 

Keywords

Bayesian estimation


graph estimation

loss functions

Markov chain Monte Carlo

network estimation 

View Abstract 3747

Co-Author(s)

Elissa Bailey
Jacob Andros, Texas A&M University

First Author

David Dahl, Brigham Young University

Presenting Author

David Dahl, Brigham Young University

Sequential Monte Carlo for probabilistic object detection in images

Object detection is a fundamental image analysis task that is relevant to many scientific disciplines. For example, astronomers detect stars and galaxies in astronomical images and biologists characterize cells and tissues in histological images. Distinguishing visually overlapping objects in images is challenging, as there is inherent ambiguity in the positions and properties of these objects. The Bayesian paradigm is well suited to this task because it provides calibrated uncertainty estimates for crowded scenes and enables scientists to incorporate prior knowledge about the imaged objects. We propose count-stratified SMC (CS-SMC), a novel Bayesian approach to object detection based on sequential Monte Carlo. Given an image, CS-SMC evaluates latent variable catalogs corresponding to various object counts and infers the posterior distribution over sets of objects. Although these sets vary in size, CS-SMC does not require transdimensional sampling, unlike existing methods for probabilistic object detection based on Markov chain Monte Carlo. We demonstrate the advantages of CS-SMC in a case study involving astronomical images of densely populated starfields. 

Keywords

sequential Monte Carlo

Bayesian inference

image analysis

object detection

astronomical images

histological images 

View Abstract 2431

Co-Author

Jeffrey Regier, University of Michigan

First Author

Timothy White, University of Michigan

Presenting Author

Timothy White, University of Michigan