Arrival and registration (welcome coffee + snacks)

Opening remarks

Keynote I

Stephan Günnemann

Graph ML for Molecules: Prediction, Generation, and Electronic Modelling

Graph Machine Learning has emerged as a powerful paradigm for modeling molecular systems: bridging the gap between computational efficiency and accuracy. It has the potential to revolutionize areas from chemistry, over materials science, to drug discovery. This talk explores the intersection of Graph ML and molecular sciences, focusing on three key areas: prediction, generation, and electronic modelling.

Benchmarking Positional Encodings for GNNs and Graph Transformers (Florian Grötschla)

Maximally Expressive Graph Neural Networks for Outerplanar Graphs (Maximilian Thiessen)

Non-Uniform Is Non-Enough (Eran Rosenbluth)

Comparing Hierarchical Network Partitions Based on Relative Entropy (Christopher Blöcker)

A Neuro-Symbolic Framework for Answering Graph Pattern Queries in Knowledge Graphs (Tamara Cucumides)

EvA: Evolutionary Approach to Attacking GNNs (Soroush H. Zargarbashi)

Towards Principled Graph Transformers (Luis Müller)

EDGE: Evaluation of Diverse Knowledge Graph Explanations (Stefan Heindorf)

Thermodynamics-informed graph machine learning for molecules and mixtures (Jan Rittig)

A Semi-Supervised Clustering Approach For Graph Learning with Neural Networks (Chester Tan)

Poster Session

Dinner

Socializing, live music by Blue Shift (Jan Borchers (p), Martin Grohe (b), Michael Krämer (dr), Christof Melcher (gtr))

Opening remarks

Keynote II

Rebekka Burkholz

Graphs as Computational or Data Structure? A Tale of Two Functions

Message passing graph neural networks (GNNs) are a powerful class of machine learning models to learn on and from graphs. By design, graphs do not only serve as data but are also utilised as computational structure. However, not all graphs are equally effective in facilitating vertex communication, often suffering from challenges like over-squashing and over-smoothing. In this talk, we explore how these issues are intertwined and propose graph rewiring strategies to mitigate their effects. Our analysis further reveals that a critical yet often overlooked factor is the limited trainability of GNNs. While techniques such as balanced initialization, dynamic rescaling, and architectural innovations can improve trainability, we show that delaying the learning of specific layers can sometimes enhance generalization, particularly in the context of homophilic tasks. Synthesizing these insights, we will propose a potential path toward multi-purpose GNN architectures and learning algorithms that disentangle the conflicting roles of graphs as data and computational structure.

Coffee break

Preventing Representational Rank Collapse by Splitting the Computational Graph (Andreas Roth)

Expressivity and Generalization: Fragment-Biases for Molecular GNNs (Niklas Kemper)

Walk-based Node Centralities for Efficient Subgraph GNNs (Fabrizio Frasca)

Exploring the Power of Graph Neural Networks in Solving Linear Optimization Problems (Chendi Qian)

Improving Subgraph-GNNs via Edge-Level Ego-Network Encodings (Vicenç Gómez)

Graph Neural Networks Do Not Always Oversmooth (Bastian Epping)

The Effectiveness of Curvature-Based Rewiring and the Role of Hyperparameters in GNNs Revisited (Floriano Tori)

A Modality agnostic graph-based framework to unify multi-modal single-cell and spatial transcriptomics data (Sikander Hayat)

Topological trajectory classification on Simplicial Complexes (Vincent Grande)

Hierarchical Graph Pooling Based on Minimum Description Length (Jan von Pichowski)

Explainable Graph Learning in Power System Applications (Sebastian Pütz)

Lunch break

Panel discussion: Graph Learning versus Computer Vision

Final remarks