Data Processing Using Machine Learning to Analyze Floor Layout Plan Design

Thanks to a growing body of research into the application of machine learning in architecture, architecture is moving into the digital realm. One obstacle is the lack of large datasets and uniform data encoding. To address this question, this thesis establishes an architecture data processing framework for using machine learning to analyze floor plan layout designs. It proves preliminarily that the framework, from data extraction to data encoding validation, has the potential to provide a large amount of data for quantitative analysis of architectural spatial quality. In this thesis, the technical aspects of this newly proposed architecture data processing framework are presented and discussed. For data extraction, graph theory is applied to extract a graph-based dataset. For feature engineering, isovist analysis is applied to convert spatial information into features. For data encoding, the graph embedding technique is used to explore a variety of data encoding methods. For data validation, the spatial features are linked to architectural qualities, specifically, five criteria for evaluating apartment rooms’ spatial quality are applied. In this framework two software programs for extracting and converting the graph-based dataset are outlined. From a proof-of-concept experiment, it could be concluded that the choice of different spatial features and data encoding methods will affect the accuracy and efficiency of floor plan analysis. And it’s necessary to select a reasonable combination according to the analysis and prediction objectives. By presenting a novel approach to architecture data processing for analyzing floor plans, this thesis opens the gate for data-driven architectural analysis and consolidates the foundation for the architecture field to enter the big data era.