Resumen
Confronted with the spatial heterogeneity of the real estate market, some traditional research has utilized geographically weighted regression (GWR) to estimate house prices. However, its predictive power still has some room to improve, and its kernel function is limited in some simple forms. Therefore, we propose a novel house price valuation model, which is combined with geographically neural network weighted regression (GNNWR) to improve the accuracy of real estate appraisal with the help of neural networks. Based on the Shenzhen house price dataset, this work conspicuously captures the variable spatial regression relationships at different regions of different variables, which GWR has difficulty realizing. Moreover, we focus on the performance of GNNWR, verify its robustness and superiority, and refine the experiment process with 10-fold cross-validation. In contrast with the ordinary least squares (OLS) model, our model achieves an improvement of about 50% on most of the metrics. Compared with the best GWR model, our thorough experiments reveal that our model improves the mean absolute error (MAE) by 13.5% and attains a decrease of the mean absolute percentage error (MAPE) by 13.0% in the evaluation on the validation dataset. It is a practical and powerful way to assess house prices, and we believe our model could be applied to other valuation problems concerning geographical data to promote the prediction accuracy of socioeconomic phenomena.