Interactive animated visualizations of probabilistic models

Taka, Evdoxia (2023) Interactive animated visualizations of probabilistic models. PhD thesis, University of Glasgow.

Abstract

Bayesian probabilistic models’ structure (determined by the mathematical relations of the model’s variables) and outputs (i.e., the posterior distributions inferred through Bayesian inference) are complex and difficult to grasp and interprete without specialized knowledge. Various visualizations of probabilistic models exist but it is very little known about whether and how they support users’ comprehension of the models. The aim of this thesis is to investigate whether adding interaction or animation to visual representations of probabilistic models help people better understand the structure of models and interprete the (causal and non-causal) relations of the variables.

This research presents a generic pipeline to transform a probabilistic model expressed in a Probabilistic Programming Language (PPL) and associated inference results into a standardized format which can then be automatically translated into an interactive probabilistic models explorer (IPME). IPME provides at-a-glance communication of a model’s structure and uncertainty, and allows interactive exploration of the multi-dimensional prior or posterior MCMC sample space. A collapsible tree-like structure represents the structure of the model in IPME. Each variable is represented by a node that presents graphically the prior or posterior distribution of the variable. Slicing on indexing dimensions or forming conjunctive restrictions on variables by interacting with the distribution visualizations is supported. Each user interaction with the explorer triggers the reestimation and visualization of the model’s uncertainty. This closed-loop exchange of responses between the user and the explorer allows the user to gain a more intuitive comprehension of the model. IPME was designed to enhance informativeness, transparency and explainability and ultimately, the potential of increasing trust in models.

This research investigates also whether adding interactive conditioning to classical scatter plot matrices that present samples from the prior distribution of probabilistic models helps users better understand the models, and if there are levels of structural detail and model designs for which it is beneficial. A user study was conducted. The analysis of the collected data showed that interactive conditioning is beneficial in cases of sophisticated model designs and the difference in response time between the interaction and static group becomes less important in higher levels of structural detail. Participants using interactive conditioning were more confident about their responses overall with the effect being stronger in tasks of lower level of detail.

This research proposes a pipeline to generate simulated probabilistic data from interven tions applied on causal structures that are expressed in PPLs using probabilistic modeling and Bayesian inference. An automatic visualization tool for visualizing the simulated probabilistic data generated by this pipeline was developed. A user study to evaluate the proposed tool was conducted. How effectively and efficiently people identify the causal model of the presented data and make decisions on interventional experiments when the uncertainty in the simulated data of interventions was presented using static, animated, or interactive visualizations was investigated. The findings suggested that participants were able to identify the causal model of the presented data either given a single intervention or by exploring various interventions. Their performance in identifying sufficient interventions was poor. Participants did not rely on the sufficient interventions to identify the causal model in the case of multi-interventional tasks. They might have relied more on combining information from multiple interventions to draw their conclusions. There were three different visual exploration strategies of the information in the scatter plot matrices which participants followed; roughly 1/3 of them relied on both the scatter and KDE plots, another 1/3 of them relied more on the scatter plots, and the last 1/3 of them relied more on the KDE plots. Those who followed the last strategy had a better performance in identifying the causal model given a specific intervention. Most participants judged the design of the visualization positively with many having mentioned that “it was informative”.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: Q Science > QA Mathematics > QA75 Electronic computers. Computer science
• Colleges/Schools: College of Science and Engineering > School of Computing Science
• Funder's Name: Engineering and Physical Sciences Research Council (EPSRC)
• Supervisor's Name: Williamson, Dr. John H.
• Date of Award: 2023
• Depositing User: Theses Team
• Unique ID: glathesis:2023-83903
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 03 Nov 2023 11:42
• Last Modified: 24 Nov 2023 10:03
• Thesis DOI: 10.5525/gla.thesis.83903
• URI: https://theses.gla.ac.uk/id/eprint/83903

Related URLs

• Article DOI
• Article DOI
• Data DOI