House Price Valuation Model Based on Geographically Neural Network Weighted Regression: The Case Study of Shenzhen, China

Zimo Wang, Yicheng Wang, Sensen Wu, Zhenhong DuIsprs Int. J. geo inf 2022 · DOI

TL;DR

This paper introduces a novel house price valuation model, Geographically Neural Network-Weighted Regression (GNNWR), designed to better capture the complex spatial heterogeneity of real estate markets. By combining the strengths of neural networks and geographically weighted regression, GNNWR achieves significantly improved prediction accuracy compared to traditional methods like OLS and GWR, offering a more robust tool for real estate appraisal.

What Problem Does This Paper Address?

Estimating house prices is crucial for economic planning and individual decision-making. Traditional methods often struggle with spatial heterogeneity, meaning that the factors influencing house prices can vary significantly across different geographic locations. For instance, proximity to transportation might be a major driver in suburban areas, while school quality is more critical in urban centers. Geographically Weighted Regression (GWR) was developed to address this by allowing regression coefficients to vary locally. However, GWR has limitations, including a restricted choice of kernel functions and potential issues with overfitting. This paper aims to overcome these limitations by proposing GNNWR, a model that leverages the learning power of neural networks to more accurately model spatial nonstationarity and improve prediction accuracy.

Methods

The core of this paper is the introduction and application of the Geographically Neural Network-Weighted Regression (GNNWR) model. GNNWR builds upon the principles of GWR by incorporating a neural network to learn a more flexible and accurate spatial weight matrix.

Here's a breakdown of the key methodological components:

  • Geographically Weighted Regression (GWR): This is the foundational technique. GWR allows regression coefficients to vary spatially, capturing local relationships. It uses a weighted least squares approach where weights are determined by the proximity of data points to a specific location.
  • Spatial Weighted Neural Network (SWNN): This is the novel component. GNNWR uses a SWNN to learn the spatial weight matrix. The SWNN takes spatial distances as input and outputs the weights, allowing for a more complex and data-driven representation of spatial relationships than traditional GWR kernel functions (like Gaussian or bi-square).
  • Neural Network Architecture: The GNNWR model employs a fully connected neural network with multiple hidden layers. Techniques like dropout and batch normalization are used to enhance generalization and prevent overfitting.
  • 10-Fold Cross-Validation: To ensure the robustness and reliability of the model, a 10-fold cross-validation approach is used. This involves dividing the data into training, validation, and testing sets to rigorously evaluate the model's performance and generalization ability.
  • Model Comparison: The performance of GNNWR is compared against Ordinary Least Squares (OLS) and GWR models using various evaluation metrics.

Datasets

The study utilizes a dataset of second-hand housing transactions in Shenzhen, China. This dataset includes:

  • Reference prices for 3595 residential quarters released by the Shenzhen Real Estate and Urban Construction Development Research Center in February 2021.
  • Geographical coordinates (latitude and longitude).
  • Endogenous variables: Building age (AB), number of parking spots (NPS), management fee (MF), green ratio (GR), and plot ratio (PR).
  • Environment-related variables: Distance from the sea (SD), quality of available public schools (QAPS), number of subway stations within a 1 km radius (NSS), and distance to the nearest subway station (DSS).

The dataset consists of 2871 complete and effective initial data records.

Key Insights and Ideas

  • Neural Networks Enhance Spatial Weighting: The core insight is that using neural networks to learn the spatial weight matrix in GWR (as in GNNWR) can significantly improve the model's ability to capture complex spatial nonstationarity compared to fixed kernel functions.
  • Beyond Simple Linearity: GNNWR's ability to learn intricate relationships means it can better model how factors like distance to the sea or proximity to schools influence house prices in nuanced ways across different parts of Shenzhen.
  • Spatial Heterogeneity is Key: The study strongly reinforces the idea that spatial heterogeneity is a critical factor in house price valuation, and models that explicitly account for it (like GWR and GNNWR) outperform global models (like OLS).
  • GNNWR Offers a "Smoother" Interpretation: Visualizations of GNNWR coefficients show a more continuous and interpretable spatial distribution compared to the more clustered and abrupt patterns seen in GWR, suggesting better handling of spatial variations.

Key Results

The GNNWR model demonstrated superior performance across various metrics:

  • Improved Accuracy: GNNWR achieved a significant reduction in prediction errors.
  • Mean Absolute Error (MAE) was improved by 13.5% compared to the best GWR model.
  • Mean Absolute Percentage Error (MAPE) decreased by 13.0% compared to GWR.
  • RMSE was approximately 13.0% lower than GWR.
  • The mean residual error of GNNWR was 62.2% lower than GWR, indicating more unbiased predictions.
  • Higher R²: GNNWR achieved a higher R² value (0.817178 on the test set) compared to GWR (0.790389) and OLS (0.432101).
  • Robustness: The 10-fold cross-validation and hypothesis testing confirmed the model's robustness and the significance of spatial nonstationarity in the Shenzhen housing market.
  • Variable Influence: The analysis revealed that while some variables like Management Fee (MF) and Sea Distance (SD) showed consistent correlations with house prices, others like Number of Parking Spots (NPS) and Building Age (AB) had more notable spatial variations in their influence when spatial heterogeneity was considered.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that while GNNWR is more accurate, the "black box" nature of neural networks can still pose challenges for interpretability compared to simpler models.
  • The paper notes that the linearity, homoscedasticity, independence, and normality properties of the error term in the linear model might not always be satisfied, suggesting potential preprocessing steps.
  • Future Work:
    • Investigate the use of the Box-Cox method for preprocessing dependent variables if the error term assumptions are not met.
  • Incorporate a wider range of independent variables to expand the model's predictive power.
  • Conduct comparative tests on different datasets or across multiple cities to establish a benchmark for housing price prediction.
  • Further explore the interpretability of the GNNWR model, potentially by analyzing the learned kernel functions more deeply.

Summary updated

BibTeX
@article{wang2022house,
  author = {Zimo Wang and Yicheng Wang and Sensen Wu and Zhenhong Du},
  title = {{House Price Valuation Model Based on Geographically Neural Network Weighted Regression: The Case Study of Shenzhen, China}},
  journal = {Isprs Int. J. geo inf},
  year = {2022},
  doi = {10.3390/ijgi11080450},
}
Figure