VF-GraphSAGE: Vertex feature-enhanced graph neural network for room classification in residential floor plans

Graph-based floor plan understanding has advanced rapidly through generative models, yet the analytical task of room-type classification — a prerequisite for layout validation and a quality signal for downstream synthesis — has received far less attention. This paper introduces VF-GraphSAGE, a graph neural network that fuses explicit room-boundary geometry with inductive graph learning. Each room is a node in a plan-level access graph described by two streams: a compact geometric descriptor and a fixed-length encoding of its raw polygon vertices, combined with GraphSAGE message passing over door-adjacency topology. We establish the first classification benchmark on the ResPlan corpus (17,107 vector floor plans; 162,755 rooms; six categories) under a strict plan-level split. VF-GraphSAGE attains 97.92 ± 0.06% accuracy and 97.93 ± 0.05% macro-F1, significantly outperforming rule-based, classic machine-learning, and GNN baselines (p < 10 -16 ). Controlled ablations attribute the gain specifically to the polygon vertex encoding. We release code, weights, and the full evaluation protocol.