ارزش‌گذاري انبوه املاك با روش يادگيري عميق و لحاظ وابستگي مكاني و زماني

Mass Real Estate Valuation Using Deep Learning Considering Spatial an‎d Temporal Dependencies

مهندسي نقشه برداري

Mass real estate valuation, as a cornerstone of urban policy-making, tax assessment, an‎d mortgage lending, has consistently attracted researchersʹ attention. However, many conventional automated valuation methods fail to simultaneously an‎d structurally incorporate the inherent spatial an‎d temporal dependencies in housing markets. This research is designed with the aim of a comparative eva‎luation of deep learning an‎d spatial model approaches in modeling these dependencies. For this purpose, a composite dataset of 5,551,648 valid residential property transactions in Englan‎d an‎d Wales from 2011 to 2019 was utilized, adjusted for inflation using the Consumer Price Index (CPI) with 2011 as the base year. Two models were implemented an‎d eva‎luated under identical conditions (identical data, stratified split based on local authority code an‎d transaction year, an‎d identical eva‎luation metrics): an optimized Deep Neural Network (DNN) architecture with 9 hidden layers that each layer has 512, 256, 128, 64, 32, 16, 8, 4, an‎d 2 neurons respectively, an‎d a Spatial Autoregressive (SAR) model with 15 significant variables an‎d a neighborhood structure of k=8. The results demonstrated that the DNN model achieved superior performance on the test set (RMSE = £46,103, R² = 0.9005) compared to the SAR model (RMSE = £73,241, R² = 0.7491); this difference equates to a 37% reduction in prediction error. Analysis of the spatial structure of residuals using Moranʹs I test revealed that spatial autocorrelation in the DNN residuals was approximately 2.5 times lower than that of the SAR model, indicating the DNNʹs superior ability to capture complex spatial structures without requiring explicit formulation. Furthermore, temporal stability analysis over the 2011–2019 period showed that the DNN model exhibited less performance fluctuation an‎d more consistent results across all years compared to the SAR model. A controlled experiment excluding spatial features confirmed their critical impact on modeling accuracy; their removal led to a 42% decrease in the R² value. These findings demonstrate that deep neural networks, despite lacking an explicit spatial structural mechanism, can model spatio-temporal structures more effectively than traditional spatial models by learning complex non-linear relationships between spatial features (such as distance to underground stations an‎d city centers). This research paves the way for employing deep learning methods in developing more transparent an‎d equitable mass valuation systems that can simultaneously account for spatial dependencies an‎d temporal dynamics.

5

158966