Prof. Dr. Tobias Friedrich


Papers accepted at STACS, AISTATS and ARES

We are happy to announce three recently accepted papers of our group members: At the 40th International Symposium on Theoretical Aspects of Computer Science (STACS) the paper Strongly Hyperbolic Unit Disk Graphs is going to be presented in Hamburg, Germany on 7-9 March. The paper introduces the class of hyperbolic unit disk graphs, which contains Euclidean unit disk graphs as well as strongly hyperbolic unit disk graphs, featuring grid-like and hierarchical network structures, respectively. The latter thus captures properties of real-world graphs like the internet. A greedy routing scheme is developed and its analysis on strongly hyperbolic unit disk graphs yields an improvement over general performance lower bounds.

Additionally, there is the paper Fast Feature Selection with Fairness Constraints at the 26th International Conference on Artificial Intelligence and Statistics (AISTATS), taking place on 25-27 April in Valencia, Spain. They received over 2000 abstract submissions this year. Of the 1689 submissions that proceeded to review, 29% were ultimately accepted to the conference. Feature Selection (FS) is a central problem in Machine Learnining, since it improves scalability and it prevents overfitting. However, common FS techniques, e.g., Forward-Stepwise Selection and Lasso-type Regularization, cannot handle complex side constraints imposed by applications such as fairness. In tthe paper, the authors address this problem, by providing the first scalable FS algorithm suitable for fair learning.

And finally, at 39th Annual Meeting & Conference of the American Real Estate Society (ARES) on 28 March - 1 April in San Antonio, USA our group will present Automated Valuation Models: Improving Model Performance by Choosing the Optimal Spatial Training Level. This paper is the outcome of the interdisciplinary Sales Comparison Approach project. The authors study the impact of the spatial training level for machine learning methods (such as random forests and neural networks) predicting the value of a property. In particular, they consider four spatial levels of a dataset of roughly 1.2 million residential properties from Germany. They observe that the right choice of spatial training level can have a major impact on the model performance, and that this impact varies across the different methods.

  • Strongly Hyperbolic Unit ... - Download
    Bläsius, Thomas; Friedrich, Tobias; Katzmann, Maximilian; Stephan, Daniel Strongly Hyperbolic Unit Disk GraphsSymposium Theoretical Aspects of Computer Science (STACS) 2023: 13:1–13:17
  • Fast Feature Selection wi... - Download
    Quinzan, Francesco; Khanna, Rajiv; Hershcovitch, Moshik; Cohen, Sarel; Waddington, Daniel G.; Friedrich, Tobias; Mahoney, Michael W. Fast Feature Selection with Fairness ConstraintsArtificial Intelligence and Statistics (AISTATS) 2023: 7800–7823
  • Automated Valuation Model... - Download
    Krämer, Bastian; Stang, Moritz; Doskoč, Vanja; Schäfers, Wolfgang; Friedrich, Tobias Automated Valuation Models: Improving Model Performance by Choosing the Optimal Spatial Training LevelAmerican Real Estate Society (ARES) 2023: 1–26