19.09.2025
From 15–19 September in Porto, RC Trust showcases research that makes AI more reliable, explainable, and usable in practice.
Photo: RC Trust
Machine learning’s flagship in Europe, ECML PKDD, brings researchers, industry, and early-career scholars together to turn advanced methods into real-world value. RC Trust joins with a clear focus: making AI dependable where it matters—healthcare, industry, and public services. Colleagues from our Data Science and Data Engineering group, Verification and Formal Guarantees of Machine Learning group (both RC Trust & Department of Computer Science) and our Causality group (RC Trust & Department of Statistics) contribute across the program, from scientific talks to workshops and the PhD Forum. Against this backdrop, our team presents two results and a workshop paper that push trustworthy AI forward—compressing complex images into interpretable structures, coordinating multiple agents with causal knowledge, and scaling anomaly detection without labels.
The highlights below include preprints for readers who want to dive deeper.
Research highlights (with preprints)
Decentralizing Multi-Agent Reinforcement Learning with Temporal Causal Information — Jan Corazza, Hadi Partovi Aria, Hyohun Kim, Daniel Neider, and Zhe Xu
Injecting temporal causal rules as prior knowledge lets multiple AI agents learn efficiently with limited communication—useful for logistics, robotics, and emergency response. The paper provides theory and case studies for faster, safer coordination.
Preprint: download here
Unsupervised Surrogate Anomaly Detection — Simon Klüttermann, Tim Katzke, Emmanuel Müller
Instead of modeling full data distributions without labels, the method learns simple “surrogate” patterns and flags deviations. A deep ensemble (DEAN) scales well and performs competitively across many datasets—promising for fraud, manufacturing, and health monitoring.
Preprint: download here
From Pixels to Graphs: Deep Graph-Level Anomaly Detection on Dermoscopic Images — Dehn Xu, Tim Katzke, Emmanuel Müller
Presented at the 22nd International Workshop on Mining and Learning with Graphs (MLG 2025), held jointly with ECML PKDD 2025. Dermoscopic images are transformed into graphs of meaningful regions so models analyze relationships—not millions of raw pixels. The pipeline reduces noise and computation while improving the detection of suspicious skin lesions.
Read more here