Enhancing the computational efficiency of the SGWR model and introducing its software implementation
The SGWR model is a local spatial regression model that functions similarly to Geographically Weighted Regression (GWR) but incorporates a spatial weight matrix derived from both geographical proximity and data attributes similarity. The integration of these two weights is controlled by a parameter...
Saved in:
| Main Authors: | , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Taylor & Francis Group
2025-06-01
|
| Series: | Annals of GIS |
| Subjects: | |
| Online Access: | https://www.tandfonline.com/doi/10.1080/19475683.2025.2523739 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The SGWR model is a local spatial regression model that functions similarly to Geographically Weighted Regression (GWR) but incorporates a spatial weight matrix derived from both geographical proximity and data attributes similarity. The integration of these two weights is controlled by a parameter (alpha), which determines their respective contributions to the regression. While the previous SGWR model outperforms GWR statistically, it remains computationally expensive, particularly during alpha optimization. In this study, we improved the model’s computational efficiency in two ways: (1) optimizing the steps used to calculate the Corrected Akaike Information Criterion (AICc) for both alpha optimization and model fitting, and (2) implementing Message Passing Interface (MPI) for parallel processing. Additionally, we developed a Python package that implements both sequential and parallel versions of the model, along with a tool to enhance accessibility. Furthermore, we incorporated a bi-square kernel function alongside gaussian kernel function, allowing for greater flexibility in spatial weighting and model adaptation to different data structures. The experiment results illustrate that the new SGWR’s computation efficiency is significantly improved, particularly for larger datasets. For instance, with the housing dataset containing 20,833 observations, the computation time decreased from 117 minutes to approximately 28 minutes when using a single CPU core, and from 56.4 minutes to just over 11 minutes with six cores for alpha optimization. The total computation time decreased from 64 minutes to 18 minutes when using 6 cores. |
|---|---|
| ISSN: | 1947-5683 1947-5691 |