GeoRDF2Vec - Spatial KGs (SKG)

GeoRDF2Vec – Learning Location-Aware Entity Representations in Knowledge Graphs

arXiv:2504.17099v1 [cs.LG] 23 Apr 2025 

Many large knowledge graphs not only encode relational information between entities but also capture the geographic geometries of some or all of these entities. Prominent knowledge graphs such as DBpedia [17], YAGO [24], and Wikidata [32] contain geographic information on entities like places and buildings. Additionally, dedicated geographic knowledge graphs, such as KnowWhereGraph [14], WorldKG [8], and OSMh3KG [3], explicitly model geographic relationships.

Exemplos de SKG

3. Böckling, M., Paulheim, H., Detzler, S.: A Planet Scale Spatial-Temporal Knowledge Graph Based On OpenStreetMap And H3 Grid. In: Geospatial Linked Data Workshop. arXiv (2024)

8. Dsouza, A., Tempelmeier, N., Yu, R., Gottschalk, S., Demidova, E.: WorldKG: A World-Scale Geographic Knowledge Graph. In: Proceedings of the 30th ACM International Conference on Information & Knowledge Management. pp. 4475–4484. ACM, Virtual Event Queensland Australia (Oct 2021). https://doi.org/10.1145/3459637.3482023

14. Janowicz, K., Hitzler, P., Li, W., et al.: Know, Know Where, KnowWhereGraph: A densely connected, cross-domain knowledge graph and geo-enrichment service stack for applications in environmental intelligence. AI Magazine 43(1), 30–39 (Mar 2022). https://doi.org/10.1002/aaai.1204

For geographical question answering, several models have been developed to generate spatially explicit KG embeddings. Unlike standard QA benchmark datasets, QA datasets that involve spatial data must account for additional constraints. One model that employs a specialized embedding approach for geographic data is TransGeo, an adaptation of TransE [19].

Tarefas com SKG: QA para Where

Mann et al. [20] address the challenge of predicting spatial links in geographic KGs, which are often sparsely interlinked. Traditional link prediction methods typically rely on existing entity relations, which may be missing in such KGs. To overcome this limitation, the authors propose two approaches:

Supervised Spatial Link Prediction (SSLP): This method leverages spatial and semantic embeddings derived from the literal values of geographic entities to predict links. 

Unsupervised Inductive Spatial Link Prediction (USLP): Unlike SSLP, this approach does not require labeled training data. Instead, it uses the haversine distance between geohash grid cell centroids to infer spatial links. 

The authors evaluate these methods using WorldKG, demonstrating that both SSLP and USLP outperform existing state-of-the-art link prediction techniques. Their results highlight the effectiveness of incorporating spatial and semantic embeddings to enhance the completeness of geographic knowledge graphs [20].

 Tarefas com SKG: Predição de links para nós espaciais. Será que poderiam ser usados para inferir contexto espacial inexistente?

In many public knowledge graphs, whether they model only spatial data or general information, spatial geometries are represented as literals within a KG. Our approach aims to leverage these geometries to model spatial relationships between individual nodes.

Novos relacionamentos entre os nós espaciais mas é baseado somente em distância

In a KG, each node can has a neighborhood. The neighborhood of a node v ∈ V is defined as H(v) := {(s, p, o) : ∃p, o|(v, p, o) ∈ E}∪{(s, p, o) : ∃s, p|(s, p, v) ∈ E}. The set of neighbors N(v) of a node v ∈ V is defined as N(v) := {v′ : ∃r|(v, r, v′) ∈ H(v) ∨ (v′, r, v) ∈ H(v)}. Geometries can be points, lines, polygons, multipoints, multilines, multipolygons and geometry collections. We define a spatial knowledge graph SKG as a knowledge graph where a part of the nodes has a geometry attached, i.e.,
SKG = (V,E, S), where S ⊆ V × G is the set of nodes that have a geometry attached, and G is the (theoretically infinite) set of all possible geometries. The subset GV ⊆ V of geographic nodes is the set of nodes which have at least one geometry, i.e., GV := {v ∈ V |∃g : (v, g) ∈ S}. We further distinguish fully spatial KGs as those KGs where every node has a geometry attached (i.e., V = GV ), and partially spatial KGs where some, but not all nodes have a geometry attached (GV ⊂ V ). The majority of common knowledge graphs, such as DBpedia or Wikidata, are partially geographic KGs.

 Definição de SKG

GeoSPARQL has been introduced as a standard for geographical representation in KGs. It defines six classes and 36 object properties. The superclass of GeoSPARQL is named SpatialObject, which represents spatial entities and includes two primary subclasses: Feature and Geometry. 

Linguagem de manipulação: extensão do SPARQL para este tipo de dados

3.2 From Spatial Distances to Edge Weights

Não são uma Algebra de Contradominio, são calculos de distância