VaidhyaMegha/football-kg

Football World Cup Knowledge Graph

16,150 nodes. 12,384 edges. 92 years of FIFA World Cup football as a graph — 30 tournaments, 1,248 matches, 3,637 goals, 10,401 players, 88 national teams and 240 stadiums.

Built with Samyama Graph. Loader and ETL: samyama-ai/football-kg.

What this is

Every FIFA World Cup from 1930 to 2022 — 22 men's tournaments (1930–2022) and 8 women's tournaments (1991–2019) — modelled as a property graph rather than a set of tables.

Because it is a graph, questions that need joins across several hops are direct: which stadiums hosted the most matches across tournaments, which players scored in more than one World Cup, how goals distribute by minute and period across 92 years.

Licence

Data	PDDL 1.0 — Open Data Commons Public Domain Dedication and Licence
Source	DataHub World Cup Datasets
Loader code	Apache-2.0, in the GitHub repo. This does not govern the data.

PDDL places the data in the public domain: no attribution requirement, no share-alike. Attribution to DataHub and to this repository is appreciated but not required.

Files

Nodes carry an id; edges reference those ids as src and tgt. Join on id to reconstruct the graph in any tool.

nodes/

File	Rows	Columns
player.csv	10,401	id, player_id, birth_date, count_tournaments, family_name, given_name, position
goal.csv	3,637	id, goal_id, minute, own_goal, penalty, period
match.csv	1,248	id, match_id, away_score, date, extra_time, home_score, name, penalty_shootout, result, stage
manager.csv	475	id, manager_id, country, family_name, given_name
stadium.csv	240	id, stadium_id, capacity, city, country, name
team.csv	88	id, team_id, code, confederation, name, region
country.csv	31	id, name
tournament.csv	30	id, tournament_id, count_teams, host_country, name, winner, year

edges/

File	Rows	Connects
scored_by.csv	3,637	Goal → Player
scored_in.csv	3,637	Goal → Match
home_team.csv	1,248	Match → Team
away_team.csv	1,248	Match → Team
in_tournament.csv	1,248	Match → Tournament
played_at.csv	1,248	Match → Stadium
from.csv	88	Team → Country
hosted_by.csv	30	Tournament → Country

Also included

football.sgsnap — the same graph as a Samyama snapshot (gzip-compressed JSON-lines). Loads into the engine in one step, without running the ETL.

Usage

As tables

from datasets import load_dataset

players = load_dataset("<ORG>/football-kg", "player")
print(players["train"].num_rows)          # 10401

As a graph

docker run -p 6379:6379 -p 8080:8080 samyama/samyama-graph
curl -X POST localhost:8080/api/tenants/default/snapshot/import \
     -F "file=@football.sgsnap"

Then query it in Cypher:

// Stadiums that hosted the most World Cup matches
MATCH (m:Match)-[:PLAYED_AT]->(s:Stadium)
RETURN s.name AS stadium, s.city AS city, count(m) AS matches
ORDER BY matches DESC LIMIT 5

// Players who scored in more than one tournament — three hops,
// the kind of question a flat table cannot answer directly.
// `given_name` is the string "not applicable" for mononyms, so match on family_name.
MATCH (p:Player)<-[:SCORED_BY]-(g:Goal)-[:SCORED_IN]->(m:Match)-[:IN_TOURNAMENT]->(t:Tournament)
WITH p, count(DISTINCT t) AS tournaments, count(g) AS goals
WHERE tournaments > 1
RETURN p.family_name AS player, p.given_name AS given, tournaments, goals
ORDER BY goals DESC LIMIT 10
// 307 players qualify. Top: Marta (17 goals, 5 tournaments), Miroslav Klose (16, 4),
// Ronaldo (15, 3), Gerd Müller (14, 2), Birgit Prinz (14, 4).

Provenance

To reproduce this graph, or to check it against the raw data, follow the chain:

1. Original source — DataHub World Cup Datasets, released under PDDL 1.0. The upstream collection provides:

tournaments.csv · teams.csv · stadiums.csv · matches.csv · players.csv · squads.csv · goals.csv · managers.csv, and optionally referees.csv · bookings.csv · substitutions.csv · penalty_kicks.csv · tournament_standings.csv · group_standings.csv

2. ETL — samyama-ai/football-kg, which maps those tables to a property graph. etl/loader.py is the entry point.

3. This dataset — exported from a Samyama snapshot taken 2026-07-01 (sgsnap format v2, engine v1.1.0), then flattened back to CSV. Node and edge totals were asserted against the snapshot header: 16,150 and 12,384, and re-verified after upload.

This build uses a subset of the upstream tables. It draws on tournaments, teams, players, matches, goals, stadiums and managers. It does not include squads.csv, referees.csv, bookings.csv, substitutions.csv or the standings tables. See Limitations for what that means in practice.

Versioning

This dataset is a git repository. Every upload is a commit, and releases are tagged so that a citation stays valid even as the card is edited.

Version	Date	Contents
v1.0	2026-08-17	Initial release — 16,150 nodes, 12,384 edges from the 2026-07-01 snapshot

Pin the tag, not main. main moves; a tag does not.

from huggingface_hub import hf_hub_download

hf_hub_download(
    "VaidhyaMegha/football-kg",
    "nodes/player.csv",
    repo_type="dataset",
    revision="v1.0",        # stable; omit this and you follow main
)

from datasets import load_dataset
load_dataset("VaidhyaMegha/football-kg", "player", revision="v1.0")

Policy for later releases: any change to the data gets a new tag (v1.1, v2.0). Corrections to this card alone do not — they land on main, and existing tags keep pointing at the data they described.

Limitations

Read this before using the graph for analysis.

• 9,045 of the 16,150 nodes (56%) have no edges. This graph connects players to matches only through goals, so a player appears connected only if they scored. 8,570 of 10,401 players are therefore isolated, and all 475 managers are isolated because no relationship type connects them to teams or tournaments.
• No squad, appearance, or lineup edges. You cannot answer "which players were in which squad" from this data. The football-kg schema defines a PLAYED_FOR edge (Player → Team, carrying tournament_id, position and shirt_number) built from squads.csv — this export does not contain it, which is why most players are isolated.
• No referees, bookings, substitutions or penalty-kick records, and no tournament or group standings, though the upstream DataHub collection and the repo's schema cover them.
• Country coverage is coarse — 31 Country nodes, several of which are confederation regions rather than states (e.g. "Africa").
• Goals carry minute, period, penalty and own_goal, but not the assisting player.
• given_name uses the literal string "not applicable", not an empty value, for players and managers known by a single name — 474 players (Marta, Ronaldo, Pelé …) and 7 managers. Concatenating given_name + family_name will produce "not applicable Marta". Filter on that string, or use family_name alone. One player has an empty birth_date.

Verified against the graph: the totals reproduce the real World Cup records — Marta 17 goals across 5 tournaments (all-time record), Miroslav Klose 16, Ronaldo 15.

Citation

PDDL places the data in the public domain, so no citation is legally required. Cite it anyway — it is what lets a reader get back to the raw data and check the work.

Cite the original source — this is where the data comes from:

DataHub World Cup Datasets. Open Data Commons Public Domain Dedication
and Licence (PDDL 1.0). https://datahub.io/collections/football

Cite this graph build — the derived graph, pinned to a revision:

Football World Cup Knowledge Graph, v1.0 (16,150 nodes, 12,384 edges).
Built with Samyama Graph. https://huggingface.co/datasets/VaidhyaMegha/football-kg
ETL: https://github.com/samyama-ai/football-kg
Source data: DataHub World Cup Datasets (PDDL 1.0),
             https://datahub.io/collections/football

BibTeX:

@misc{football_kg_samyama,
  title        = {Football World Cup Knowledge Graph},
  howpublished = {\url{https://huggingface.co/datasets/VaidhyaMegha/football-kg}},
  version      = {v1.0},
  note         = {Derived from DataHub World Cup Datasets (PDDL 1.0),
                  \url{https://datahub.io/collections/football}.
                  ETL: \url{https://github.com/samyama-ai/football-kg}},
  year         = {2026}
}

The three pieces

Piece	Where
Code — ETL, schema, loaders	github.com/samyama-ai/football-kg
Data — this dataset, plus the raw source it derives from	here, and datahub.io/collections/football
Graph — importable snapshot	football.sgsnap in this repository