Benchmarking methods for classifying space functions and access elements in multi-unit apartment buildings

Machine learning (ML), graph deep learning (GDL), natural language processing (NLP), generative, and image deep learning (IDL) methods are promising for automating space function and space access element classification for building analysis. Benchmarking these five methods is currently infeasible primarily due to a lack of datasets with diverse data representation formats. This paper introduces SFC-A68, a dataset derived from 275 apartments in 13 countries, to address this. The dataset was used to develop three state-of-the-art models for each of the five methods. Benchmarking predictive performance resulted in a GDL model, HGAT, achieving the highest weighted average F1-Score of 95.0%, surpassing other models by 11.0% or more. Moreover, GDL models required less pre-processing and no post-processing, fewer trainable parameters, and shorter training times than NLP, Generative, and IDL models. ML models were less accurate than IDL models but required fewer trainable parameters, shorter training times, and least pre- and no post-processing. • Five classification methods were benchmarked for space function classification. • A tabular, graph, and image dataset was created for 275 apartments in 13 countries. • GDL model HGAT had highest accuracy, least parameters and fastest training. • NLP and Generative models had lowest accuracy. • ML models were less accurate than IDL models.