emergenCITY scientist develops T-Rex selector, an efficient framework that enables fast variable selection in large, high-dimensional data sets.
Finding missing people under rubble after an earthquake is a major challenge, even for rescue robots. The situation is similar in medicine: finding genes that are responsible for specific diseases is complicated. Sometimes it’s like looking for a needle in a haystack. For applications such as these, emergenCITY scientist Jasin Machkour has developed an efficient method, the T-Rex selector. A paper on the T-Rex Selector has now been published in Signal Processing, one of the most renowned journals for signal processing, and another paper has already been accepted for publication.
Under the title “The terminating-random experiments selector: Fast high-dimensional variable selection with false discovery rate control”, the authors Jasin Machkour, Michael Muma, and Daniel P. Palomar present a method that controls a user-defined false discovery rate (FDR) in large-scale, high-dimensional data and simultaneously maximizes the number of selected variables. The T-Rex algorithm can deliver reproducible results in significantly less time than benchmark methods. While these require five hours of computation time for variable selection problems with 5,000 variables, the T-Rex selector manages five million variables in the same time.
“This is relevant for the emergenCITY rescue robot Scout, for example,” Jasin Machkour, research assistant in the Robust Data Science Group at TU Darmstadt, says. “The radar data can be used to find the right directions, in which missing persons are located, more quickly while reducing false alarms.”
In general, the T-Rex selector could be of interest for various applications, including medical studies. For example, there is already a collaboration with Charité Berlin to investigate whether the methods are suitable for achieving the same results as conventional methods with less data, i.e. with fewer study participants.
The open-source software package also developed by the researchers has already sparked great interest. It has already been downloaded 16,000 times since it was made publicly available.
In another article, which will soon be published in the Journal Signal Processing, as well as two conference papers, which will be presented in April at the International Conference on Acoustics, Speech, and Signal Processing (IEEE ICASSP), the largest conference for researchers in the field of signal processing, the scientists present further methodological and algorithmic results as well as software solutions on the topic, including an extension to complex-valued data.
Jasin Machkour, Michael Muma, Daniel P. Palomar, The terminating-random experiments selector: Fast high-dimensional variable selection with false discovery rate control, Signal Processing, Band 231, 2025, 109894, ISSN 0165-1684, https://doi.org.
Robust Data Science Group at TU Darmstadt: www.etit.tu-darmstadt.de/rds/index.en.jsp
T-Rex-Software Package: https://cran.r-project.org
