Anchor-Enhanced Geographical Entity Representation Learning.

Area of Science:

• Geoinformatics
• Machine Learning
• Spatial Data Science

Background:

• Geographical entity representation learning (GERL) embeds geographical entities into vector spaces for geographical intelligence applications.
• Existing GERL models struggle with the unbalanced spatial distribution of geographical entities, leading to inadequate representations.
• Uniform treatment of all entities in prior GERL models fails to capture nuanced spatial relationships.

Purpose of the Study:

• To propose an anchor-enhanced GERL (AE-GERL) model to improve the accuracy of geographical entity embeddings.
• To address the challenge of unbalanced spatial distributions in geographical data.
• To enhance the utility of geographical entities in diverse geographical intelligence applications.

Main Methods:

• Developed an anchor selection algorithm to identify key informative entities based on spatial distribution and types.
• Constructed an anchor-enhanced graph to explicitly link anchors with non-anchor entities.
• Employed a graph neural network (GNN) based model for anchor-to-nonanchor node learning to impute missing information.

Main Results:

• AE-GERL significantly outperforms baseline models across four diverse datasets.
• The model demonstrates superior performance in both sparse and dense geographical entity distribution scenarios.
• Experimental results validate the effectiveness of using anchors to improve entity representations.

Conclusions:

• The proposed AE-GERL model offers a methodological advancement for embedding geographical entities.
• This approach provides an effective strategy for enhancing geographical intelligence applications.
• The study highlights the benefit of incorporating informative anchors in graph-based learning for spatial data.