A Geographically Weighted Regression Method Based on Regional Attribute Similarity

As geographic data continue to expand in scale and complexity, classical geographically weighted regression (GWR) faces notable limitations in capturing local abrupt changes and directional spatial patterns, such as sharp housing-price discontinuities across rivers or the directional spread of infectious diseases. The main challenge is its inability to effectively represent the similarity relationships among surrounding environmental attributes, which limits its adaptability to the diverse variability of complex spatial structures. To overcome this limitation, a GWR approach based on regional attribute similarity is proposed, in which a single-attribute neighborhood set is defined for each covariate within the bandwidth. Localized attribute weights are further derived to reflect the spatially varying importance of covariates, thereby guiding the contribution of different attributes to neighborhood similarity measurement. The second-order structure of each set is extracted as a semivariance vector using the mark variogram and normalized into a distributional form. Differences between neighborhood distributions are quantified using the earth mover’s distance to construct pairwise attribute-similarity weights, which are then integrated with geographic similarity weights to produce hybrid regression weights that simultaneously capture spatial proximity and regional structural concordance. The experimental results demonstrate that the proposed method outperforms test approaches in terms of regression accuracy, model stability, and spatial pattern explanatory power, thereby offering an effective framework for the quantitative analysis of complex geographic phenomena.