XSRGCNN: Learning deep spatial heterogeneity via explainable spatial regression graph convolutional neural networks
Topics:
Keywords: Spatial regression; Deep spatial heterogeneity; Spatial varying coefficients; Explainable GeoAI; Graph convolutional neural networks; Local model
Abstract Type: Paper Abstract
Authors:
Di Zhu, University of Minnesota
,
,
,
,
,
,
,
,
,
Abstract
It is challenging to incorporate parametric heterogeneity in a spatial regression model. Misspecification of the model can lead to inevitably incorrect interpretation of spatial varying coefficients in empirical studies. Our motivation in this work is to bypass the statistical assumptions in current spatial regression models and learn the spatial varying knowledge that is more representative of the true spatial heterogeneity. To achieve this, we design a deep explainable spatial regression model based on the graph convolutional neural network (XSRGCNN), the goal is to learn and evaluate a deep yet explainable representation of the spatial associations within the spatial regression setting. The representativeness of learned deep spatial varying coefficients is evaluated via a simple diagnostic test, which examines the power of determinants of those local coefficients on the target variable.
XSRGCNN: Learning deep spatial heterogeneity via explainable spatial regression graph convolutional neural networks
Category
Paper Abstract