News - Data Science Chair

In our paper "Enhancing Sequential Next-Item Prediction through Modelling Non-Item Pages" we investigate the utility of non-item pages in transformer based recommender models for next item prediction and show how information from pages like category pages or search results can be leveraged.

In our paper "Liquor-HGNN: A heterogeneous graph neural network for leakage detection in water distribution networks "" we introduce the Liquor-HGNN model, a novel approach for detecting and localizing leaks in drinking water distribution networks (DWDNs) through the utilization of heterogeneous graph learning.

In our paper "Can Neural Networks Distinguish High-school Level Mathematical Concepts?" we evaluate the ability of various neural networks to classify the relationship between two mathematical expressions. For example, two expressions can be equivalent or one could be a derivative of the other. We show that neural networks are able to learn the underlying mathematical patterns and outperform rule-based systems.

Our paper "Higher-Order DeepTrails: Unified Approach to *Trails" presents an alternative approach using autoregressive language models to the Bayesian methods HypTrails, MixedTrails and SubTrails.

In this technical report, we detail the process of setting an automated speech detection pipeline on a low-power device. We expand work done previously at our chair and deploy a trained Siamese network together with a k-NN-classifier.

We have successfully acquired funding for a three year DFG project in cooperation with the Chair for Computer Philology and History of Contemporary German Literature (Prof. Fotis Jannidis). The goal of the project is to improve the capabilities of Language Models to model characters in literary texts - an area where even current models often still struggle to provide a good representation of the text's content, the story. If you're interested in joining, please contact us!

We show that architectures from the machine learning domain of continual learning can help language models to procedurally learn facts about the world.

We design a multi-agent simulation that can produce company data that also contains hidden fraud cases.

We rigorously evaluate commonly used approaches that explain the decisions of network attack detection systems.

In our work we propose a method for learning explicit representations of PDEs from non-grid data