GeoDFNet: a point-of-interest classification algorithm with dual fusion of geospatial local neighborhood features

Current Point of Interest (POI) classification models predominantly depend on textual data for feature modeling, often failing to resolve ambiguities in POI naming conventions. To overcome this limitation, we propose a Geospatial local neighborhood Dual Fusion Network (GeoDFNet), which synergizes multimodal features through a hierarchical fusion framework. By leveraging geographic similarity principles, GeoDFNet first constructs POI-centric local neighborhoods by encoding spatial relationships and aggregating surrounding geographic features via graph attention networks (GAT). In parallel, a Transformer encoder extracts latent semantic representations from textual metadata. The model employs a multi-head attention mechanism coupled with a dual-phase fusion strategy to dynamically calibrate the contributions of text and geospatial features. Experimental results on three real-world datasets (Shanghai POI, Beijing AutoNavi, and Guangdong OpenStreetMap) demonstrate that GeoDFNet achieves significantly higher classification accuracy compared to baseline models. Notably, on the Shanghai dataset, GeoDFNet attained an overall accuracy of 98.60%, substantially outperforming all textual baselines ( e.g ., Text Convolutional Neural Networks (TextCNN): 91.13%, Text Recurrent Neural Networks (TextRNN): 83.54%, Transformer: 86.64%).These experimental results confirm that the proposed model achieves robust performance and effectively mitigates the ambiguity issue in POI names.