Past seminars

• 11/11/2025 at 12:00 (BM.2.26) – Lukas Zierahn 

  • Title: Best Arm Identification for Bandits with Shifting Means

  • Abstract: We study an interactive testing environment in which the learner chooses one of K distributions to sample each timestep with the goal of finding the distribution with the highest expected mean with high probability in the least number of overall samples (the best-arm identification setting with known confidence). We coin the Shifting Means setting, a non-stationary setting where an offset is added to all samples which can change fully adversarially between timesteps and demonstrate how the known approaches of the Track-and-Stop algorithm and the Generalized Likelihood Ratio Test (GLRT) fail for Shifting Means. We propose generalized Importance Weights for Shifting Means, which we show to be delta-correct and sample efficient. We also present a lower bound and evaluate our results empirically.

• 25/11/2025 at 14:00 – Hidde Fokkema

  • Title: Performativity and Risks of Algorithmic Recourse

  •  Abstract: Algorithmic recourse has emerged as a popular explainability approach for making machine learning decisions more transparent and actionable, offering individuals guidance on how to overturn unfavorable outcomes. In this talk, I will present recent theoretical work that critically examines the consequences of recourse. First, we show that providing recourse can decrease the accuracy of the model, by encouraging users to move into regions of greater predictive uncertainty.  Second, we explore how the performative nature of recourse can render recourse invalid upon model retraining, because the recourse recommendations will induce a distribution shift. We identify conditions under which this performative invalidity arises. Together, these findings challenge the assumption that recourse is always beneficial, and we will discuss some situations where recourse could be beneficial and when not. This talk is based on work with Damien Garreau, Tim van Erven, Gunnar König, Timo Freiesleben, Celestine Mendler-Dünner and Ulrike von Luxburg.

• 28/10/2025 at 12:00 (CE.0.18) – Tianjiao Yan  

  • Title: Alternative representations of the Aalen-Johansen estimator of the state occupation probabilities in multi-state models

  •  Abstract: Multi-state models are a flexible framework for analysing processes that evolve through different states over time, with applications in medicine, demography, and economics, among others. A key quantity of interest is the state occupation probability, commonly estimated by the Aalen–Johansen estimator. In this work we provide alternative representations of this estimator by extending both the redistribution-to-the-right algorithm and the inverse probability of censoring weighted empirical average representation to the general multi-state setting. We show that these formulations are mathematically equivalent to the Aalen–Johansen estimator, offering new interpretative insights and practical understanding.