In this master's thesis, we will investigate the relationship between actionable recourse, explainable AI (XAI), causality, and reliability under model and data changes. The student will study whether recourse recommendations remain valid and stable when models are retrained or the underlying data distribution shifts.
One possible direction is to compare conventional counterfactual explanations with causality-aware recourse, where recommended actions respect relationships between variables and distinguish directly actionable features from immutable or indirectly changeable ones. The project may also investigate the alignment between explanations and recourse: do the features identified as important by an explanation actually correspond to the changes recommended by the recourse method?
A central technical component of the thesis will be the development of a simulation framework for systematically studying recourse under controlled model and distribution shifts. The simulator can represent changing populations, retraining of decision models, different recourse strategies, and causal or feasibility constraints. It can then be used to evaluate when recommendations remain stable and successful and when they become unreliable. Depending on the student's interests, the thesis may investigate questions such as:
- How stable are actionable recourse recommendations under model retraining and distribution shift?
- Do different but similarly accurate models recommend different actions to the same individual?
- Can explanation stability help identify unreliable recourse?
- Does causality-aware recourse produce more realistic and robust recommendations?
- Can we develop an explanation–recourse alignment or reliability score indicating when a recommendation should not be trusted automatically?
- How can a simulation framework be designed to evaluate recourse reliability under controlled changes over time?
The project combines machine learning, explainable AI, actionable recourse, causality, and simulation, and provides opportunities for both methodological research and development of a reusable experimental artifact.