XAI for understanding deep neural networks in regression
This thesis focuses on explaining deep neural network regression models. While much of the work in explainable AI (XAI) has centred on classification models, regression problems are equally significant. Furthermore, machine learning models used for regression share the need for powerful XAI methods that promote transparency, reliability, and robustness. However, key differences, such as the continuous nature of regression outputs and the focus on quantitative relationships between variables, demand tailored explanatory approaches that set regression apart from classification in the XAI domain. We begin by examining the challenges these differences pose when applying XAI attribution methods to explain complex, non-linear regression models. In particular, we identify the completeness property of attribution methods as crucial for ensuring that the model’s output unit is fully accounted for in the attributions. Additionally, we emphasize the importance of customizing explanations to address the specific explanatory needs posed by the user in the regression context.
To address the latter, we introduce two approaches, retraining and restructuring, which allow the incorporation of a user-defined reference value relative to which the explanation is contextualized. We further extend these solutions into a broader framework called XpertAI, which improves usability in combination with state-of-the-art XAI attribution methods and provides additional insights into how these two approaches relate. The effectiveness of these contextualization techniques is demonstrated through extensive quantitative and qualitative benchmark experiments.
Finally, we apply our methods to a real-world problem, showing how the regression-specific challenges can be tackled when explaining wind turbine power-curve models. Using a newly developed validation framework, we ensure that the machine learning models employ physically reasonable strategies and that the XAI attributions provide insights that are meaningful to domain experts. We then demonstrate how properly contextualized attributions can be effectively used for turbine monitoring.
In conclusion, the findings of this thesis underscore the critical importance of using XAI techniques that are tailored to the specific problem, context, and domain. Our contributions mark a step forward in enhancing the understanding of deep neural networks in regression, moving closer to more effective and context-aware explanations.