Presenting a Predictive Model of Residential Facade Preferences UsingMachine Learning Case Study: Tehran

The issue of urban facade preferences for users is one of the most important issues in the field of urban design. The answer to this question has been considered by researchers in the form of various objective and subjective methods. This study intends to use machine learning method as a predictable approach to evaluate the preferences and also desirability of urban facades for users. Therefore, the aim of the research is to design a predictive model that its output is the level of user preferences of residential facades in Tehran. According to the expected output, the data provided to the model consists of residential facade image. Due to the necessity of standard data in the machine learning process, residential facade images submitted to the Tehran City Facade and Landscape Commission in the years 2016 to 2019 have been used. Out of the original 800 images, 278 images were chosen in selection process. The input of this predictive model is images along with features. The features considered in this research have been obtained using the approach of visual preferences and image processing. The issue of whether the physical characteristics related to the visual preferences approach and the statistical characteristics obtained with the image processing technique both have an effect on the level of visual preference was tested with machine learning and the results showed that the use of both the feature provide better results. Since the supervised machine learning method has been used, it was necessary to present the labels to the machine. Therefore the number of preferences were carried out through an online questionnaire by users (218) in four categories of low preferences (0-25%), medium (50-26%), good (51-75%) and very good (100-76%). By selecting the models and determining the amount of 80 to 20 as the training to test data volume, the learning process was carried out and then using the confusion matrix, the validity of the models used in machine learning was tested. Also, to ensure the predictability of the machine, at the end, some new facades which were neither training nor test data were presented to the machine and the degree of predictability of their visual preference was checked by the machine and with the result of the survey. Based on the results, three algorithms of support vector machine, decision tree and random forest with 100% accuracy and X-G-Boost method with 97% accuracy have performed best. Based on the results, the importance of the influence of elements on users' preferences, includes the minimum distance between windows, the ratio of transparent to opaque surface in the facade, the presence of gardens in the balcony, the variety of materials, the maximum distance between windows, the number of openings, the length of the balcony, the number of balconies, the number of floors, the variety of colors, the decorations used on the roof, the type of roof lines (continuous, discontinuous), the number of entrances, and the ratio of the height to the length of the building.