Can graph machine learning read a real, complex high-rise the way it reads a clean dataset?
We try to both, understand how the building works as a circulation network, and predict what each room is from layout structure alone
THE DATA SET
The dataset and the schema — the benchmark we force our tower into.
From Floor to Graph
the process
KIFAF TOWERS
Atkins Engineering, Dubai, UAE
The building, the levels we slice, and the hand-labelled ground truth.
SCOPE OF SPATIAL STUDY
GF/ PODIUM +3rd/4th Floor- typical
GRAPH CONSTRUCTION
Space syntax in graph form — turning the feel of a plan into measurable numbers.
INITIAL SETUP: Edges, Nodes, Grid
INITIAL SETUP: Stairs/Lifts
INITIAL SETUP: Stairs/Lifts
INITIAL SETUP: Combined Building Graph
SPATIAL ANALYSIS
Space syntax in graph form — turning the feel of a plan into measurable numbers.
SPATIAL GRAPH ANALYSIS: Shortest Path
CROSS FLOOR SHORTEST PATH
SPATIAL GRAPH ANALYSIS: Closeness Centrality
SPATIAL GRAPH ANALYSIS: Betweenness Centrality
SPATIAL GRAPH ANALYSIS: Community
SPATIAL GRAPH ANALYSIS: Community –Areas
SPATIAL GRAPH ANALYSIS: Degree centralities
SPATIAL GRAPH ANALYSIS: Degree centralities
SPATIAL GRAPH ANALYSIS: Visibility Heatmap
NODE CLASSIFICATION
Train a GNN to name each room from its features and its neighbours — then test on held-out rooms
BEHIND THE SCENE: rhino to TopologicPy
BUILDING THE CELL COMPLEX: Three Identical Floors
PERFORMING ADJACENCY GRAPH: Three Identical Levels
ACCESSIBILITY GRAPH
BUILDING THE CELL COMPLEX: One Floor
BUILDING THE CELL COMPLEX: One Floor
PERFORMING ADJACENCY GRAPH: Connected Levels
ROOM PREDICTION
ROOM PREDICTION SUMMARY-3rd floor
ROOM PREDICTION SUMMARY-4th Floor
ROOM PREDICTION COMMUN ERROR
CONCLUSIONS
Can graph ML read a real, complex high-rise like a clean dataset? Mostly yes
We read Kifaf Towers two ways:
As a circulation network — closeness, betweenness and isovists pinpoint the corridors and cores that move people through each floor. Useful for wayfinding and evacuation.
As a room graph — a GNN names rooms from layout alone (~87% on L04, 95% on L03). A messy Rhino model can be auto-labelled like curated research data.
