LaM-SLidE: Latent Space Modeling of Spatial Dynamical Systems via Linked Entities

• DOI: 10.48550/arxiv.2502.12128
• Publication Date: 2025-02-17
• Authors: Florian Sestak; Artur P. Toshev; Andreas Fürst; Günter Klambauer; Andreas Mayr; J. Brandstetter
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; Artificial Intelligence (cs.AI); Computer Science - Artificial Intelligence; Machine Learning (cs.LG)

Generative models are spearheading recent progress in deep learning, showing strong promise for trajectory sampling dynamical systems as well. However, while latent space modeling paradigms have transformed image and video generation, similar approaches more difficult most systems. Such -- from chemical molecule structures to collective human behavior described by interactions of entities, making them inherently linked connectivity patterns the traceability entities over time. Our approach, LaM-SLidE (Latent Space Modeling Spatial Dynamical Systems via Linked Entities), combines advantages graph neural networks, i.e., across time-steps, with efficiency scalability advances where pre-trained encoder decoder frozen enable generative space. The core idea is introduce identifier representations (IDs) allow retrieval entity properties, e.g., coordinates, system thus enables traceability. Experimentally, different domains, we show that performs favorably terms speed, accuracy, generalizability. (Code available at https://github.com/ml-jku/LaM-SLidE)