Diepgrond, Denny (2020) Can prediction explanations be trusted? On the evaluation of interpretable machine learning methods. Master's Thesis / Essay, Artificial Intelligence.
|
Text
mAI_2020_DiepgrondD.pdf Download (2MB) | Preview |
|
![[img]](https://fse.studenttheses.ub.rug.nl/style/images/fileicons/text.png) |
Text
toestemming.pdf Restricted to Registered users only Download (96kB) |
Abstract
Explanations of machine learning models try to help a user decide whether to trust its predictions. While performance metrics for machine learning models have been well established and are important to today’s AI successes, performance metrics for explanations of model predictions are as yet less well investigated. The question we want to answer in this work goes a step further than trusting predictions: Can we trust the explanations of machine learning predictions? The contribution of this work is twofold. First a theoretical framework to evaluate methods for interpretable machine learning is established based on regulatory requirements and social explanation theory. Secondly the framework is applied (in part quantitatively and in part qualitatively) to evaluate two state-of-the-art explanation methods (LIME and Kernel SHAP) on synthetic datasets with known explanatory structure. The predefined data distributions have served as a ground truth that made objective evaluation of the explanations possible. Moreover, an intuitive assessment has been proposed that serves as a first step towards a general evaluation of explanation models that are in production. Our results suggest that evaluating explanations of model predictions should become as integrated in the field of machine learning as evaluating performance of the models themselves. The analysis and methods in this work are a step in that direction.
| Item Type: | Thesis (Master's Thesis / Essay) |
|---|---|
| Supervisor name: | Verheij, H.B. |
| Degree programme: | Artificial Intelligence |
| Thesis type: | Master's Thesis / Essay |
| Language: | English |
| Date Deposited: | 02 Jun 2020 10:44 |
| Last Modified: | 02 Jun 2020 10:44 |
| URI: | https://fse.studenttheses.ub.rug.nl/id/eprint/21985 |
Actions (login required)
 |
View Item |