Title:  Understanding the algorithmic bias of graph neural networks

Abstract:  There is growing interest in learning algorithmic reasoning directly from data. In this talk, we survey recent results on the ability of graph neural networks to express, learn, and generalize algorithmic procedures from finite training data. First, we present a theoretical framework for characterizing when a graph neural network can learn an algorithm in a supervised setting and provably generalize beyond the training distribution. Second, we describe an unsupervised framework for learning to solve the uniform facility location problem with a constant-factor approximation guarantee.

Bio: Christopher Morris studied Computer Science at TU Dortmund University, Germany. In 2019, after a short stint at Stanford University, he completed his Ph.D. at TU Dortmund, focusing on machine learning for graph and relational data. He then spent one year as a postdoctoral fellow at Polytechnique Montréal in the Department of Mathematical and Industrial Engineering, followed by a postdoctoral position in the Computer Science Department at McGill University and as a member of Mila – Quebec AI Institute. In June 2022, he joined RWTH Aachen University, Germany, as a tenure-track assistant professor in the Computer Science Department and was promoted to full professor in  December 2025.

Abstract: Combinatorial optimization is fundamental to decision-making across science, engineering, and industry, with Mixed-Integer Linear Programming (MILP) serving as one of its most powerful general-purpose frameworks. In this talk, I will explore recent advances in deep learning for MILP, from foundational neural architectures to real-world industrial deployment. We present an attention-based general neural backbone that improves MILP representation learning beyond conventional GNNs, followed by applications in real-time game matchmaking and large-scale industrial production planning. These works demonstrate how learning can significantly accelerate MILP solving and expand its practical impact on complex optimization problems.

Bio: Dr. Wen Song received his Bachelor’s and Master’s degrees from Shandong University, China, and the PhD degree from Nanyang Technological University, Singapore. He is currently an Associate Professor with Shandong University, China. His research interests include artificial intelligence, combinatorial optimization, planning and scheduling. His work on learning driven scheduling received the 2024 IEEE TII outstanding paper reward. He served as an Area Chair/Senior PC for top conferences such as ICML, ICLR and NeurIPS. He is a Senior Member of IEEE.