Toward Robust Urban Region Representation Learning Through Hierarchical Modeling for Modifiable Areal Unit Problem Mitigation

Region representation learning encodes urban regions into a unified latent vector space, facilitating various urban intelligence tasks. However, existing methods often rely on predetermined scales and shapes, overlooking the modifiable areal unit problem (MAUP), which limits their ability to capture fine-grained spatial relationships. Although hierarchical modeling offers a potential solution, its application to urban contexts is challenged by irregular boundaries, global dependencies, and heterogeneous data sources. To address these issues, we propose HierRE, a hierarchical region representation learning framework that captures multiscale features while accommodating diverse spatial configurations. HierRE introduces a hexagonal patchifying strategy to handle irregular boundaries by dividing regions into consistent spatial units, and applies a hexagonal positional encoding (HexPE) based on cube coordinates to preserve spatial relationships. It also employs a hierarchical architecture to model global dependencies and multiscale interactions, and integrates a multiview feature fusion mechanism to combine heterogeneous urban data such as points-of-interest (POIs) and remote sensing (RS) imagery. Extensive experiments demonstrate that HierRE outperforms state-of-the-art baselines across four socioeconomic indicator prediction tasks and one land use classification task. Visualization studies further highlight HierRE’s interpretability and its ability to generate robust and meaningful region representations.