Characterizing the Performance of Counterfactual and Correlation Guidance via Dataset Perturbations

Loading...
Thumbnail Image
Date
2025
Journal Title
Journal ISSN
Volume Title
Publisher
The Eurographics Association
Abstract
Guidance methods are often employed in visual analytics systems to help users navigate complex datasets and discover meaningful insights. Guidance based on correlation is a common method that can steer users towards closely related variables. However, recent work has shown that guidance based on counterfactual subsets can more effectively capture and surface causal relationships. In this work we further explore these guidance methods by characterizing their performance by systematically introducing perturbations in both the data points generated from a ground truth causal graph, and the causal relationships in the graph itself. Our results indicate that while both guidance types exhibit similar sensitivity to global data point perturbations, counterfactual guidance can better capture perturbations affecting only a single dimension, and more effectively reflect changes in causal link strengths, indicating an improved ability to capture narrow data changes and causal relationships.
Description

        
@inproceedings{
10.2312:evp.20251124
, booktitle = {
EuroVis 2025 - Posters
}, editor = {
Diehl, Alexandra
and
Kucher, Kostiantyn
and
Médoc, Nicolas
}, title = {{
Characterizing the Performance of Counterfactual and Correlation Guidance via Dataset Perturbations
}}, author = {
Wang, Arran Zeyu
and
Borland, David
and
Gotz, David
}, year = {
2025
}, publisher = {
The Eurographics Association
}, ISBN = {
978-3-03868-286-8
}, DOI = {
10.2312/evp.20251124
} }
Citation