Ensuring the quality of modern methods of machine learning with regard to predictive quality, robustness, impartiality and interpretability is often a challenge in practical applications. Our research focuses on:
- Comprehensive generation and selection of features for complex learning procedures (e.g. for use in online scenarios and taking into account different uncertainty factors).
- Robust predictive and explanatory mechanisms, taking into account various fairness definitions.
- Techniques for quantifying and improving the quality of (training) data.
Open PhD Position
If you are interested in explainable Machine Learning & Data Science consider to apply to our group by sending your CV, your most recent transcript, and a writing sample to: email@example.com. We are looking forward to hearing from you. More information follows soon.
Paper @ KDD 2020
"Leveraging Model Inherent Variable Importance for Stable Online Feature Selection" by Johannes Haug, Martin Pawelczyk, Klaus Broelemann and Gjergji Kasneci has been accepted at KDD 2020.
Paper @ UAI 2020
On Learning Invariant Counterfactual Explanations under Predictive Multiplicity by Martin Pawelczyk, Klaus Broelemann and Gjergji Kasneci was accepted at UAI 2020.
Best-Paper Award @ Symposium on Eye Tracking Research and Applications (ETRA), 2020
A MinHash approach for fast scanpath classification by David Geisler, Nora Castner, Gjergji Kasneci and Enkelejda Kasneci was accepcted at ETRA (2020)
Paper @ Archives of Data Science, Series A
PLAY: A Profiled Linear Weighting Scheme for Understanding the Influence of Input Variables on the Output of a Deep Artificial Neural Network by Torsten Dietl, Gjergji Kasneci, Johannes Fürnkranz and Eneldo Loza Mencía was accepted at Archives of Data Science, Series A.
Paper @ WWW 2020
"Learning model agnostic actionable counterfactual explanations for tabular data" by Martin Pawelczyk, Klaus Broelemann und Gjergji Kasneci was accepted at WWW 2020. The paper presents a new method to generate counterfactual explanations.
Paper @WIRE's Data Mining and Knowledge Discovery Journal accepcted
Ein einleitender Survey "Bias in Data-driven AI Systems - An Introductory Survey" was accepted at WIRE's Data Mining and Knowledge Discovery Journal. This is our first multidisciplinary joint publication in the context of the EU project "NoBias": https://nobias-project.eu/index.php/partners/.
Paper @ AAAI 2020
'Training decision trees as replacement for convolution layers' by Wolfgang Fuhl, Gjergji Kasneci, Wolfgang Rosenstiel und Enkelejda Kasneci was accepted @ AAAI 2020: The work suggests to compute convolutions using decision trees.
Paper @ NeurIPS 2019 (HCML workshop)
"Towards User Empowerment" by Martin Pawelczyk, Johannes Haug, Klaus Broelemann and Gjergji Kasneci was accepted at NeurIPS' 'Human-Centric Machine Learning' workshop. The paper describes shortly a new way to to generate counterfactuals for heterogeneous data. Moreover, we also suggest how the quality of counterfactual suggestions can be evaluated.
Data Science Reading Group
For the first time in WS19/20, we are organizing a reading group on various topics from the field of machine learning and data science. DSAR team members alternately present interesting papers from their research areas. The first reading group will take place on Tuesday, 22.10.19, at 11 am on Sand 14 in room C207, and afterwards once every two weeks at the same time. A list of the topics will be continuously updated on our research page. All those interested are welcome to attend.
Paper @ ICANN 2019
"CancelOut: A Layer for Feature Selection in Deep Neural Networks", a paper by Vadim Borisov, Johannes Haug and Gjergji Kasneci has been accepted at ICANN 2019. This work presents a new effective and yet simple layer for feature selection in deep neural networks.
Paper @ IJCAI 2019
"A Gradient-Based Split Criterion for Highly Accurate and Transparent Model Trees", a paper by Klaus Broelemann and Gjergji Kasneci was accepted at IJCAI 2019. The work presents a new split-criterion to boost interpretability of powerful tree based algorithms.