Random Visibility Graph とGraph Neural Networkを用いた間取りの特徴量表現

抄録

Floor plans are widely used for representing residential spaces. This study aims to propose a feature representation method for floor plans, namely unstructured image data, using a Random Visibility Graph and Graph Neural Network. In this method, the feature representation is performed using Invariant Information Clustering. To prepare a dataset, the Random Visibility Graph is designed to derive multiple graphs from a single floor plan effectively. The obtained features formed shape-oriented clusters, rather than ordinary variables such as area. In addition, the similarity of features corresponds to the similarity of plans, including slight differences in the connectivity of rooms.