AC-SGB: Autism Centre Spatial Graph Benchmark; A Labelled Graph Dataset for Autism-Friendly Spatial Layout Design

This dataset provides 17,839 labelled spatial layout graphs for autism-friendly educational environments, generated by a hybrid Genetic Algorithm–Simulated Annealing (GA-SA) framework across 263 benchmark scenarios spanning cohorts of 18–50 individuals with autism. Each layout is represented as a weighted, undirected graph, with nodes encoding distinct spaces (minimum/maximum area, minimum width, a malleability-rigidity score, and external-wall requirements) and edges encoding required or candidate adjacencies weighted by adjacency strength. Every graph carries a continuous fitness label (0.0–1.0), derived from a composite fitness function combining topological similarity, geometric similarity, and a transition-area penalty against a scenario-specific optimal configuration. The dataset totals 755,768 nodes and 2,718,826 edges across the 17,839 graphs, with node counts per graph ranging from 2–65 and edge counts from 1–763. Graphs are grouped across five similarity-score bands (very low to optimal) to ensure balanced coverage of the fitness range rather than concentration around high-fitness solutions. The dataset was benchmarked against six predictive models: Linear Regression, MLP, and Random Forest (non-graph baselines), and GCN, GINE, and GraphConv (message-passing GNNs), for the task of predicting a graph's fitness score from its structure and attributes. Full methodology, benchmark results, and an out-of-distribution generalisation test (on unseen 70–200-occupant scenarios) are reported in the accompanying paper.