Floor-Plan Room Labelling and Generative Layout Using Graph Neural Networks

A floor plan is a graph: rooms are nodes; doors, shared walls and windows are typed edges. On 16,317 plans of the 2025 ResPlan data set, GraphSAGE (macro-F1 0.9890) and RGCN (0.9880) name rooms clearly better than a Random Forest (0.9641) and decision-tree rules (0.8761); the two networks tie. A small graph-conditioned network with a repair step then places one rectangle per room inside a boundary: door contact 0.925 against 0.563 for a fitted treemap, but 16 % of the floor stays empty and 37 % of rooms fall below National Building Code 2016 minimum sizes (fitted treemap 8 %, real plans 3 %). A web application labels plans, generates layouts, exports IFC and OBJ, and keeps an architect's reviews. A teaching prototype, not a design tool: a metric win does not mean a usable plan.

Category: School of Architecture and Planning

Architecture & Planning Project AP05 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.