From Spaces to Graphs: Spatial Data Representation
Speaker: Seyran Khademi
Abstract: Architectural and urban spatial data are conventionally represented through geometry—coordinates, dimensions, and visual fidelity. Yet, when the objective shifts from drawing space to reasoning about it, geometry becomes secondary to topology, where relationships, adjacency, connectivity, and circulation patterns hold the primary semantic meaning. This talk introduces a topological modeling perspective for spatial datasets, employing graph representations and graph neural networks to distill and interpret the latent structure of spatial organization beyond traditional geometric descriptions.
[ Slides ]
Floor Plan Similarity as Graph Comparison Problem
Speaker: Casper van Engelenburg
Abstract: When graphs encode spatial layouts, how can we compare them in a way that is both expressive and fast? Many methods for computing the similarity between two graphs rely on computationally expensive cross-graph message passing, which drastically limits the scalability of such methods. I will present an alternative viewpoint: learn graph representations independently and compute similarity only at the end using a differentiable graph kernel on the node embeddings. This preserves structural sensitivity while making comparison far more efficient, offering a practical route to scalable similarity learning of (spatial) graphs.
[ Slides ]