Statistical learning has achieved remarkable progress thanks to large datasets and data‑driven protocols, yet it still struggles to generalize beyond observed distributions and often fails under distribution shift. Causal models target the underlying mechanisms of change rather than surface correlations, enabling robust reasoning under external interventions. This allows systems to learn from data, simulate hypothetical scenarios, adapt to new environments, and support better decisions.
The thesis is divided into two parts. The first part studies algorithmic recourse—identifying changes an individual can make to overturn an unfavorable decision. We propose a more realistic cost function than the conventional framework, reformulate the problem in causal terms, and incorporate a temporal dimension into policy evaluation to capture delayed effects of actions.
The second part focuses on dynamical systems, extending causal interventions and effects to discrete and continuous stochastic processes. We show that the equilibria of these processes and their shifts after interventions can be linked to Structural Causal Models (SCMs), which serve as compressed representations of the underlying temporal dynamics. Because real‑world actions often take time to manifest, temporal causal models are highly relevant.
Both parts share the practical goal of enhancing the utility, robustness, and trustworthiness of models by ensuring they realistically represent the phenomena they aim to capture.
Review: By integrating causal reasoning into algorithmic recourse and dynamical systems, the work offers a more interpretable and temporally aware framework, providing solid theoretical and methodological foundations for decision‑making under distribution shift.