Geo-SAM: a structure-attribute match method for geographic data prediction

Spatial prediction aims to estimate values at unobserved locations by leveraging spatial correlations and feature dependencies. While graph neural networks (GNNs) have shown strong performance in this domain by encoding spatial proximity through graph construction and embedding node attributes, they often struggle with data that are sparse, heterogeneous, or complex. As such, we proposed the Structure-Attribute Matching (Geo-SAM) model, a novel framework that decoupled and realigned spatial and attribute information through dual-path encoding and statistical alignment. Geo-SAM employed a GNN-based structural encoder that captured spatial correlations and a neural network (NN)-based attribute encoder that processed non-spatial features. Their latent representations were aligned using Maximum Mean Discrepancy (MMD) to ensure coherence between spatial and attribute representations. To mitigate the over-smoothing, a known limitation of deep GNNs, residual connections were integrated into both encoders to retain discriminative signals. The aligned latent representations were subsequently fused and decoded to generate predictions. An adaptive loss-balancing mechanism dynamically adjusted the weight between reconstruction and alignment losses, facilitating stable optimization across varying data conditions. Experiments on synthetic and real-world datasets demonstrated that Geo-SAM consistently outperforms representative baselines, delivering state-of-the-art accuracy and robustness across diverse spatial prediction scenarios.