Distributed Heterogeneity Learning for Generalized Partially Linear Models with Spatially Varying Coefficients.

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 flexibility and parsimony in spatial regression.

Main Methods:

• Designed a distributed heterogeneity learning (DHL) method using bivariate spline smoothing.
• Implemented a scalable and communication-efficient algorithm for DHL.
• Provided theoretical guarantees for the DHL framework.

Main Results:

• The DHL method offers a simple, scalable, and communication-efficient implementation.
• Achieved almost linear speedup in computational performance.
• DHL constant coefficient estimators are asymptotically normal.
• DHL spline estimators achieve the same convergence rate as global estimators.

Conclusions:

• The proposed generalized models and DHL method effectively address spatial heterogeneity and large data challenges.
• DHL provides a computationally efficient and theoretically sound approach to spatial analysis.
• The method demonstrates practical utility through simulations and real-world data analysis.