SF Emergency Services Knowledge Graph

Results — Knowledge graph vs flat table

Multi-agency escalation

+79.6%Accuracy

KG achieves 100% accuracy vs 83.9% flat. The graph traverses Response → Unit → Agency to know exactly which agencies responded.

agency_countis_multi_agency

Resolution time

+4.3%RMSE improvement

MAE drops from 24.75 to 23.94 minutes. Graph-derived response times and location history improve prediction.

avg_response_timelocation_prior_count

Severity classification

−2.2%accuracy delta

Flat model slightly wins at 86.9% accuracy. Severity signal already exists in priority and response count.

signal already in flat features

Key insight: The knowledge graph's value is not universal — it shines when the prediction depends on relationships between entities that don't exist in a flat table. The real value is infrastructural: the same graph powers all three models, answers ad-hoc SPARQL queries, and grows without schema changes.

Knowledge graph stack

Predictive models

user-defined

Constrained logistic regression + random forest. Three targets: severity, multi-agency, resolution time.

SeverityMulti-agency?Resolution time

↑

Enrichment

user-defined

Added via APIs using location + timestamp.

Weather (NOAA)Census / ACSZoningEvents

↑

Inference

Derived via SHACL rules and inverse paths. Same CAD number links all agencies.

Incident → responsesCross-agency alignmentLocation historyisMultiAgency

↑

Foundation

published

Ontology + instance data. All datasets share the same URIs.

owl:Class defssh:NodeShapeSHACL validationFire calls ~700KPolice dispatchEMS

Property reference

Class	Object properties (→ target)	Key datatype properties
Incident	`hasResponse → Response · hasCallType → CallType · hasLocation → Location · hasSeverity → SeverityLevel`	`cadNumber · receivedTimestamp · priorityCode · responseCount · totalResolutionMinutes`
Response	`responseTo → Incident · hasUnit → Unit`	`dispatchTimestamp · onSceneTimestamp · responseTimeMinutes`
Unit	`belongsToAgency → Agency`	`unitId · unitType`
Agency	—	`title`
CallType	`realizationOf → IncidentConcept · hasAgency → Agency`	—
IncidentConcept	—	`description`
Location	`inNeighborhood → Neighborhood · inStationArea → StationArea`	`address · latitude · longitude · zipCode`
PredictionSet	`hasPrediction → Prediction`	`modelName · featureSet · accuracy`
Prediction	`predictsIncident → Incident`	`predictedProbability · predictedClass`