Spatiotemporal Knowledge Graph (STKG)

Updated 1 May 2026

Spatiotemporal Knowledge Graphs are formal representations that enrich traditional graphs by linking entities with explicit time intervals and spatial regions.
They combine static, temporal, and spatial paradigms to support diverse applications such as urban mobility, power outage analysis, and robot navigation.
Modern STKGs leverage embedding techniques and graph neural methods to boost prediction accuracy, query efficiency, and dynamic event processing.

A Spatiotemporal Knowledge Graph (STKG) is a formal knowledge representation structure that generalizes the classical knowledge graph paradigm by explicitly linking entities and relations to both temporal and spatial context. Each edge or fact in an STKG represents not only a semantic relationship between two entities but also the spatial region and time interval during which this relationship holds. The addition of spatiotemporal semantics enables integrated modeling, reasoning, and analytics over complex systems and event trajectories across scientific, industrial, urban, and environmental domains.

1. Formal Foundations and Canonical Structure

Let $E$ be a finite set of entities (nodes), $R$ a set of relation types (predicates), $T$ a temporal domain (typically in $\mathbb{Z}$ for discrete time, $\mathbb{R}$ for continuous), and $S$ a spatial domain (e.g., planar coordinates, regions, or polygons). The core structure is a multi-relational labeled graph

$G_\mathrm{STKG} = (E, R, T, S, F)$

where $F \subseteq E \times R \times E \times T \times S$ is a set of spatiotemporal facts. Each fact $f = (s, p, o, t, \ell)$ asserts that subject $s$ and object $R$ 0 are linked by predicate $R$ 1 during time $R$ 2 and at spatial context $R$ 3 or spatial region. Temporal attributes may be instants, intervals, or durations; spatial attributes may be coordinates, named regions, grid cells, or polygons.

Alternative formalisms, such as those used in applied STKGs, package facts as RDF triples with explicit spatial and temporal annotations: $R$ 4 with $R$ 5 (spatial domain), $R$ 6 (temporal domain) (Frakes et al., 30 Jul 2025), or extend the entity set and relations to include temporal and spatial entities as nodes (Zhao et al., 2024, Wang et al., 2021).

Tensor representations are standard:

$R$ 7

This structure supports general algorithmic and embedding-based analytics.

2. Modeling Dimensions and Taxonomy

STKGs have evolved from three parent paradigms: static knowledge graphs, temporal graphs, and spatial graphs—each contributing distinct modeling idioms (Plamper et al., 18 Dec 2025).

Static-graph-derived STKGs: Add time and spatial attributes to classical KG facts.
Temporal-graph-derived STKGs: Begin with edge dynamics, then spatially enrich nodes and relations.
Spatial-graph-derived STKGs: Start from spatial proximity or geometric graphs, then interleave event or relation time.
STGNN-inspired STKGs: Employ time-varying node/edge features atop a fixed or dynamic graph structure.

Modeling axes include:

Edge semantics: Directed/undirected; explicit (physical connection) or inferred (mobility, similarity).
Temporal annotation: Node-annotated (time as node property), edge-annotated (validity/activation time per edge), snapshot-annotated (graph $R$ 8 per $R$ 9).
Spatial annotation: Node- or edge-based spatial labels; multi-layer designs linking nonspatial and spatial entities; spatial context as region, coordinate, or cell.
Combined semantics: Node and edge attributes may simultaneously track spatiotemporal validity, with higher-order “events” (split, merge, transfer) encoded as edge types or subgraphs.

3. Construction, Ontology, and Data Integration Pipelines

Construction of an STKG proceeds through a pipeline designed to integrate and align multimodal data sources, develop an expressive ontology, and generate the graph representation suitable for efficient querying and analytics.

Ontology design typically extends existing vocabularies (e.g., GeoSPARQL, DBpedia, Media Resources OWL) with domain-specific classes (e.g., OutageRecord, NTLImage, OutageMap in GeoOutageKG (Frakes et al., 30 Jul 2025)), supporting both coarse (e.g., county) and fine (e.g., grid cell, pixel) spatial scales and time granularities (Frakes et al., 30 Jul 2025).

Integration methodology:

Spatial and temporal data are harmonized using region identifiers (e.g., FIPS codes, geohashes) and ISO 8601 datetimes or intervals.
Data from different modalities (e.g., time series, imagery, event records) are aligned via ontology mappings and cross-reference properties (e.g., “derivedFrom,” “hasObservationTime”).
Provenance and FAIR principles inform identifier schemes and data publication endpoints.

Pipeline stages:

Data acquisition and preprocessing
Curation, spatial/temporal masking, and filtering
Ontology mapping and IRI assignment
RDF/graph triple generation with explicit $T$ 0, $T$ 1 mappings
Publish/query: GraphDB and SPARQL/GeoSPARQL endpoint exposure, with support for subgraph or multi-resolution queries (Frakes et al., 30 Jul 2025, Zhou et al., 2024)

4. Representation Learning and Inference

The embedding of spatiotemporal facts for predictive and analytical tasks is central to STKGs.

Complex Embedding Models: Entities, relations, time, and location are embedded in $T$ 2 and scored via multi-way products. For example, STComplEx uses

$T$ 3

where $T$ 4 and “ $T$ 5” denotes elementwise product, generalizing ComplEx and TComplEx to spatiotemporal contexts (Dai et al., 2024).

Graph Neural Methods: Multi-relational GNNs or temporal GCNs propagate spatiotemporal signals through the graph, with message passing that integrates, e.g., region, time-slot, and distance embeddings, as in STKG-based recommendation or urban mobility prediction (Zhao et al., 2024, Wang et al., 2021).
Influence-based and Dynamic Embeddings: Explicit dynamic “in” and “out” node embeddings parameterized by time, fused with spatial influence coefficients determined via observed patterns, form the core of interpretable and scalable STKGs such as SSTKG (Yang et al., 2024). Optimization alternates embedding parameter learning with influence weight adaptation.
Community Detection and Motif Discovery: STKGs with weighted, temporally and spatially annotated edges can be input to graph partitioning or community detection (e.g., Louvain algorithm with weighted adjacency) to extract clusters manifesting joint spatial and temporal coherence, a methodology that demonstrates improved precision for human mobility analytics (Li et al., 2024).

5. Query, Analytics, and Visualization

Rich spatiotemporal annotations enable advanced patterns of graph querying and visualization.

Querying

SPARQL and GeoSPARQL permit retrieval conditioned on spatial region and time, as well as joins across modular data classes (e.g., aggregating time-series and imagery for disaster analysis (Frakes et al., 30 Jul 2025)).
Subgraph retrieval for reasoning uses pattern templates matching spatial constraints (bounding-box, region, proximity) and temporal intervals.
Semantically-aware question answering pipelines leverage STKG embeddings aligned to transformer LMs, as in STCQA (Dai et al., 2024).

Visualization

Multi-view platforms (e.g., GeoViz (Zhou et al., 2024)) integrate tree, net, and map representations, allowing cross-filtered exploration along spatial, temporal, and semantic axes. Knowledge maps and nets provide spatial/temporal “brushing,” concept hierarchy views, and semantic similarity nets, supporting scalable and dynamic STKG exploration.

Pattern Matching and Event Processing

Windowed, time-aggregated STKG variants (e.g., VEKG-TAG (Yadav et al., 2020)) enable complex event processing, with CEP-rule pattern matching over spatial and temporal relations, yielding low-latency analytics on compressed graphs.

6. Application Domains and Empirical Results

STKGs are foundational in several applied domains:

Domain	Key Entities	Spatial Model	Temporal Model	Reference
Urban Mobility	Users, POIs, Categories	Polygon, Point	Discretized bins	(Wang et al., 2021)
Power Outage Analysis	Outage records, counties, NTL	County, Pixel	15min/daily intervals	(Frakes et al., 30 Jul 2025)
Human Activity from CDR	Users, stays, grid cells	Grid (queen)	10min time-slots	(Li et al., 2024)
Video Event Detection	Objects, relations	Frame, region	Per-frame, window	(Yadav et al., 2020)
Robot Navigation	Objects, spatial relations	Bbox, 3D position	Frame, intervals	(Mdfaa et al., 1 Oct 2025)

Empirical results consistently show that STKG-based methods deliver improved accuracy in prediction (e.g., +5% in top-1 mobility prediction vs. RNN baselines (Wang et al., 2021)), superior temporal consistency and spatial precision in activity location detection (Li et al., 2024), and substantial gains in knowledge graph completion tasks when time and location are jointly modeled in embeddings (Dai et al., 2024). Efficiency benchmarks demonstrate sub-second or millisecond scale for pattern matching with optimized graph aggregation (Yadav et al., 2020), and scalable exploration for 10M+ triple STKGs in disaster analytics (Frakes et al., 30 Jul 2025).

7. Challenges, Modeling Guidelines, and Future Directions

Key open challenges in STKG research include heterogeneous modeling assumptions, lack of unified frameworks, and difficulty in benchmarking and provenance tracking (Plamper et al., 18 Dec 2025). Guidelines for robust STKG design recommend:

Edge formation: Choose directed/undirected and explicit/inferred relations based on domain semantics.
Temporal annotation: Prefer edge-level or graph-level (snapshot) annotation for evolution, node-level for static attributes.
Spatial annotation: Node-level for static locations; edge/graph layered for flows, events, or region transitions.
Semantic provenance and alignment: Adopt standard vocabularies (e.g., GeoSPARQL), versioned IRIs, and FAIR data principles.
Incremental update and historicity: Support for evolving graphs with traceable updates.

Unresolved research themes include the development of methods for automatic construction from raw and unstructured data (including LLM-based extraction), integration of uncertainty and fuzzy semantics, incremental and streaming updates, interpretability and explainability of deep spatiotemporal embeddings, and scalable visualization for ultra-large and dynamically evolving STKGs (Plamper et al., 18 Dec 2025).

References:

(Frakes et al., 30 Jul 2025) GeoOutageKG: A Multimodal Geospatiotemporal Knowledge Graph for Multiresolution Power Outage Analysis
(Plamper et al., 18 Dec 2025) A Survey on Spatio-Temporal Knowledge Graph Models
(Zhao et al., 2024) Spatial-Temporal Knowledge Distillation for Takeaway Recommendation
(Dai et al., 2024) Question Answering Over Spatio-Temporal Knowledge Graph
(Li et al., 2024) A Spatiotemporal Knowledge Graph-based Method for Identifying Individual Activity Locations from Mobile Phone Data
(Wang et al., 2021) Spatio-Temporal Urban Knowledge Graph Enabled Mobility Prediction
(Yang et al., 2024) SSTKG: Simple Spatio-Temporal Knowledge Graph for Interpretable and Versatile Dynamic Information Embedding
(Zhou et al., 2024) GeoViz: A Multi-View Visualization Platform for Spatio-temporal Knowledge Graph
(Yadav et al., 2020) Knowledge Graph Driven Approach to Represent Video Streams for Spatiotemporal Event Pattern Matching in Complex Event Processing
(Mdfaa et al., 1 Oct 2025) VL-KnG: Visual Scene Understanding for Navigation Goal Identification using Spatiotemporal Knowledge Graphs
(Zhao et al., 21 Apr 2026) STK-Adapter: Incorporating Evolving Graph and Event Chain for Temporal Knowledge Graph Extrapolation