Comparative Evaluation of Tensor-based Data Representations for Deep Learning Methods in Architecture

This paper presents an extended evaluation of tensor-based representations of graph-based architectural room configurations. This experiment is a continuation of examination of recognition of semantic architectural features by contemporary standard deep learning methods. The main aim of this evaluation is to investigate how the deep learning models trained using the relation tensors as data representation means perform on data not available in the training dataset. Using a straightforward classification task, stepwise modifications of the original training dataset and manually created spatial configurations were fed into the models to measure their prediction quality. We hypothesized that the modifications that influence the class label will not decrease this quality, however, this was not confirmed and most likely the latent non-class defining features make up the class for the model. Under specific circumstances, the prediction quality still remained high for the winning relation tensor type.