|
Title
|
|
|
|
Counterfactual explanations for real-world applications
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| This dissertation covers a comprehensive study of counterfactual explanations, emphasizing their importance in providing interpretability to complex machine learning (ML) algorithms, with a focus on how to deploy them in real-world scenarios. We first introduce NICE, a novel algorithm for tabular data that specifically takes into account algorithmic requirements that are often overlooked. NICE exploits information from the training data to speed up the search process of finding an explanation. Our extensive benchmarking study indicates that NICE’s explanations have desirable properties compared to the current state-of-the-art. Next, we study the implementation of NICE within the risk management system at the Belgian Customs Agency. Our research identified four stakeholders: data scientists, targeting officers, domain experts, and decision subjects. For each of the stakeholders, we investigate how they can benefit from local explanation methods and what their requirements are regarding the form of these explanations. Broadening the scope of our research beyond tabular data, we introduce SEDC and SEDC-T, two counterfactual algorithms for image classification. We experiment with various segmentation methods to find meaningful counterfactual explanations. Finally, this work tackles the ethical issue of disagreement among counterfactual explanations, revealing how their diversity can potentially be exploited to fairwash ML models by hiding sensitive features. Our study finds alarmingly high disagreement levels between the tested methods. Furthermore, we provide recommendations to avoid the disagreement problem. |
| |
|
Language
|
|
|
|
English
| |
|
Publication
|
|
|
|
Antwerpen
:
University of Antwerp, Faculty of Business and Economics
,
2023
| |
|
Volume/pages
|
|
|
|
xiv, 164 p.
| |
|
Note
|
|
|
|
:
Martens, David [Supervisor]
:
Peeters, Bruno [Supervisor]
| |
|