A Bayesian Approach to Multiview Learning

• Activity Number: 23 - Bayesian Methods and Approaches in Big Data Analysis
• Type: Topic-Contributed
• Date/Time: Sunday, August 8, 2021 : 1:30 PM to 3:20 PM
• Sponsor: International Society for Bayesian Analysis (ISBA)
• Author(s): Lasse Vuursteen* and Botond Szabo and Timo van der Poel and Wouter van Loon and Marjolein Fokkema and Mark de Rooij
• Companies: Leiden University and Vrije Universiteit Amsterdam and Leiden University and Leiden Institute of Psychology and Leiden University and Leiden University
• Keywords: Multi-view learning; Bayesian methods; Uncertainty quantification

Abstract

In multi-view learning, one considers prediction based on different sources of information called views. Each view constitutes the observations of the same label Y_i but with a different set of covariates. When one has multiple views concerning the same prediction problem, how does one find out which view is most predictive? And how should predictions based on separate views be combined?

For example, say that for each patient in a biomedical study, one can obtain separately different types of medical scans. Finding out which type of scan(s) are most valuable and how the predictions of these scans should be combined is valuable when obtaining these scans is costly.

We study the performance of Bayesian methods in a variety of simulated multi-view settings such as linear- and logistic regression. We compare the performance to established frequentist methods. In particular, we investigate the potential of Bayesian methods in quantifying the degree to which a view has predictive value.