The MaPS seminar is a hybrid event; held from 16.00 - 17.00 hrs in room ZI-4070 and via TEAMS.
Feel free to reach out if you want to present your work in an upcoming session!
Send us an email!
The MaPhS organizers,
Maike de Jongh, Patryk Rygiel and Lars Schroeder
upcoming dates:
2026:
- 14 January 2026
Speaker: Lars Schroeder
Title : node2vec random walks: Regular graphs and recurrence
Abstract: To understand the behavior of random walks, aspects like the stationary distribution, recurrence properties, diffusivity and more are studied. In this talk, we focus on tuneable random walks, called node2vec random walks, that come from the popular algorithm node2vec which is used for network embedding. The transition probabilities of the random walks depend on the previous visited node and on the triangles that contain the current and the previous node. Even though the algorithm is widely used in practice, mathematical properties of node2vec random walks almost have not been investigated. We consider these random walks on regular graphs and present results about the stationary distribution by going to a higher-order space and check under which conditions the random walks are recurrent.
Speaker: Floor van Maarschalkerwaart
Title : Perturbation-Aware Distributionally Robust Optimization for Inverse Problems
Abstract: In this talk, we will introduce “perturbation-aware Distributionally Robust Optimization (DRO)”: a flexible framework for robust reconstruction in inverse problems, building on classical distributionally robust optimization techniques. We account for uncertainty in the data (due to noise, imperfect forward models, or limited samples) by optimizing against worst-case perturbations within an ambiguity set. This set is defined using a Wasserstein ball and a prescribed class of perturbations K, which allows us to target different types of robustness: from input and output noise to conditional distributions Y|X, where we will focus on the latter. This set K also allows us to incorporate our physical knowledge of the forward model into the ambiguity set. We will show how this perturbation-aware DRO formulation leads to a general min-max problem and find and evaluate its dual form. Using this dual form, we show that perturbation-aware DRO yields tighter bounds and adaptive regularization, when compared to DRO in the full joint probability space. To illustrate the approach, we will share some examples to demonstrate how perturbation-aware DRO can produce reconstructions that are robust across a variety of noise models without retraining.