Datasets:

DropTheHQ
/

dropthe-knowledge-graph

license: mit
task_categories:
  - graph-ml
  - feature-extraction
language:
  - en
size_categories:
  - 1M<n<10M
tags:
  - knowledge-graph
  - graph-neural-networks
  - entity-linking
  - relationship-extraction
pretty_name: DropThe Entity Relationship Graph

DropThe Entity Relationship Graph

A large-scale entity relationship dataset containing 2.9 million typed, directional connections between 1.8 million entities spanning entertainment, media, finance, and technology. Extracted from the DropThe knowledge graph.

Dataset Description

While most open knowledge graphs focus on encyclopedic facts (Wikidata) or narrow domains (MovieLens for ratings), this dataset captures operational relationships -- the connections that actually matter for building recommendation systems, search engines, and analytical tools. Every edge is typed, directional, and timestamped.

Link Types

Link Type	Count	Example
`acted_in`	820,000+	Florence Pugh -> Oppenheimer
`directed`	96,000+	Christopher Nolan -> Oppenheimer
`produced`	145,000+	Emma Thomas -> Oppenheimer
`works_at`	230,000+	Tim Cook -> Apple
`founded`	18,000+	Jensen Huang -> NVIDIA
`distributed_by`	52,000+	Oppenheimer -> Universal Pictures
`related_to`	1,200,000+	Bitcoin -> Ethereum (thematic)
`traded_on`	340,000+	Bitcoin -> Binance
`parent_of`	46,000+	Alphabet -> Google

Graph Statistics

Nodes: 1,828,455 entities across 5 types (movies, series, people, companies, cryptocurrencies)
Edges: 2,947,733 typed directional links
Average degree: 3.2 connections per entity
Largest connected component: Covers 94% of all entities
Diameter: 8 hops (entertainment subgraph)

Data Format

Each edge record contains:

source_id: Entity ID of the origin node
target_id: Entity ID of the destination node
link_type: Typed relationship label (see table above)
weight: Confidence score (0.0-1.0) based on source reliability
created_at: Timestamp of link creation
source_table: Whether the entity comes from entities or geo_entities

Intended Use

This graph dataset is particularly suited for:

Graph neural networks -- Train GCN, GAT, or GraphSAGE models for link prediction and node classification
Knowledge graph completion -- Predict missing edges using TransE, RotatE, or other embedding methods
Recommendation systems -- Build multi-hop recommendation paths (actor -> movie -> director -> other movies)
Community detection -- Identify clusters in the entertainment industry (studios, talent agencies, franchise ecosystems)
Temporal analysis -- Study how industry relationships evolve over time using edge timestamps

Methodology

Links are extracted through a combination of structured API ingestion (TMDB, Wikidata, CoinGecko), automated entity resolution using the DropThe alias system (80,000+ aliases), and a bidirectional auto-linker that ensures graph consistency. The full pipeline is described on DropThe.

The enrichment process runs on a local PostgreSQL instance with validation gates that check referential integrity, duplicate detection, and type conformity before any edges are promoted to the production graph on DropThe.

Sample

The included sample_edges.csv contains 20 representative edges across multiple relationship types, demonstrating the schema and data format.

License

MIT License.

Citation

@dataset{dropthe_graph_2026,
  title={DropThe Entity Relationship Graph},
  author={DropThe},
  year={2026},
  url={https://dropthe.org}
}