Distributed Heterogeneity Learning for Generalized Partially Linear Models with Spatially Varying Coefficients
Shan Yu1, Guannan Wang2, Li Wang3
1Department of Statistics, University of Virginia, Charlottesville, VA 22904.
Related Experiment Videos
View abstract on PubMed
Summary
This study introduces a new method for analyzing spatial data, balancing model complexity and efficiency. The distributed heterogeneity learning (DHL) method effectively handles large datasets while maintaining accuracy in spatial regression.
Area of Science:
- Spatial statistics
- Econometrics
- Environmental science
Background:
- Spatial heterogeneity is crucial in various scientific fields.
- Spatially varying coefficient models address heterogeneity but reduce parsimony.
- Large spatial datasets pose computational challenges.
Purpose of the Study:
- Develop generalized partially linear spatially varying coefficient models.
- Introduce a novel distributed heterogeneity learning (DHL) method for large datasets.
- Balance model flexibility and parsimony while improving scalability.
Main Methods:
- Generalized partially linear spatially varying coefficient models.
- Distributed heterogeneity learning (DHL) using bivariate spline smoothing.
- Scalable, communication-efficient algorithm design.
Main Results:
- DHL achieves near-linear speedup for large spatial datasets.
- Theoretical guarantees for DHL: asymptotic normality and optimal convergence rates.
- DHL demonstrates effectiveness in simulations and real-world U.S. loan data.
Conclusions:
- The proposed DHL method offers a flexible, parsimonious, and scalable approach to spatial regression.
- DHL effectively handles large-scale spatial data challenges.
- Rigorous theoretical support validates the DHL framework's performance.