Two new preprints from our group
I am very happy to announce two new preprints from our group!
In the first one, my colleague Lorenz Hilfiker develops (strictly) proper scoring rules (resistance and diffusion score) for node selection tasks on static graphs. He combines his extensive knowledge of kernels and kernel scores and his expertise in network science to derive scoring rules on a domain (graphs) in which not much work has been done before. He then goes on to show how his two scores elegantly relate to classical scoring rules such as Brier or RPS, and demonstrates their practical usefulness for epidemic source detection and hierarchical classification for the ImageNet challenge.
In the second preprint (which went online shortly after the first one), my colleague Nathan Cincera Brack put out his first PhD paper (yay!). Nathan is the first to publish a thorough review of most of the (relevant) source detection methods on temporal networks that have been proposed since 2010. In order to do that, he adjusted all methods so that they apply to the same problem formulation. He then compared them theoretically and in terms of computational complexity, and, finally, performed extensive evaluations. Somewhat surprisingly, he finds that two methods outperform all others across the board: a Deep Learning-based method and a (loopy) belief propagation approach.
I congratulate both of them on their amazing research and look forward to the feedback from the community.
Both works have been generously funded by SNSF, grant number 228527.
The preprints are available on arXiv:
Hilfiker, L., Cincera Brack, N., & Sterchi, M. (2026). Strictly proper scoring rules for node selection tasks on graphs. arXiv. https://arxiv.org/abs/2611.03274
Cincera Brack, N., Sterchi, M., Hilfiker, L., & Bernstein, A. (2026). Finding Patient Zero in Temporal Contact Networks: A Comparative Study. arXiv. https://arxiv.org/abs/2611.23175