luxury-lakehouse/football2vec-player-embeddings

football2vec Player Embeddings

Pre-computed player embedding vectors from the football2vec v2 transformer model — ready to use without loading model weights. Covers ~87,000 per-match vectors, ~8,950 career vectors, and season-level aggregates across professional soccer competitions. V2 uses a 192-dim transformer encoder with adversarial team debiasing (Ganin GRL) to prevent team identity from confounding player style representations.

Part of the (Right! Luxury!) Lakehouse soccer analytics platform.

Quick Start

from datasets import load_dataset
import numpy as np

ds = load_dataset("luxury-lakehouse/football2vec-player-embeddings")
df = ds["train"].to_pandas()

# Extract behavioral vectors as a NumPy matrix
vectors = np.array(df["behavioral_vector"].tolist())
print(f"{vectors.shape[0]} players, {vectors.shape[1]}-dim embeddings")

Explore interactively: Soccer Analytics App

What Are These Embeddings?

Each embedding is composed of two complementary vectors trained or derived from open event data:

• Behavioral vector (192-dimensional): A transformer encoder embedding with adversarial team debiasing (Ganin GRL). Trained on tokenized SPADL action streams (23-type vocabulary), it captures how a player plays — their movement patterns, decision sequences, and positional tendencies — while removing team-specific confounds. V1 (32-dim Doc2Vec) is retained as a baseline.
• Statistical vector (13-dimensional, may be NULL): Per-90 statistics z-score normalized within position group (GK, Def, Mid, Fwd). This position-aware normalization prevents goalkeeper contamination in similarity search and provides fairer cross-position comparisons.

For model architecture details, training methodology, and the full vocabulary, see the companion model repositories: luxury-lakehouse/football2vec-v2 (v2 transformer) and luxury-lakehouse/football2vec-statsbomb-wyscout (v1 Doc2Vec baseline).

Configs

career (default) — ~8,950 rows

One row per player. The behavioral vector is the element-wise mean of all per-match embeddings across the player's career. Useful for overall style profiling and cross-era similarity search.

season

One row per player per competition-season. Captures style evolution across seasons and enables within-season similarity comparisons.

per_match — ~87,000 rows

One row per player per match. The most granular config; suitable for match-level clustering and fine-grained style analysis.

Data Fields

Career Config (fct_player_embeddings_career)

Column	Type	Description
canonical_player_id	string	Unified player identifier (see dim_players)
behavioral_vector	array<double>	192-dim embedding (element-wise mean across career)
stat_vector	array<double>	13-dim z-score normalized per-90 stats (may be NULL)
total_matches	bigint	Number of matches contributing to this career embedding
data_sources	array<string>	Data providers contributing (e.g., ["statsbomb", "wyscout"])

Season Config (fct_player_embeddings_season)

Column	Type	Description
embedding_season_id	string	Surrogate key
canonical_player_id	string	Unified player identifier
competition_id	int	Competition identifier
season_id	int	Season identifier
behavioral_vector	array<double>	192-dim embedding (season-level mean across matches)
stat_vector	array<double>	13-dim stat vector (may be NULL)
matches_in_sample	bigint	Number of matches contributing in this competition-season
data_sources	array<string>	Data providers contributing

Per-Match Config (fct_player_embeddings)

Column	Type	Description
embedding_id	string	Surrogate key
canonical_player_id	string	Unified player identifier
match_id	string	Match identifier
data_source	string	Data provider (statsbomb or wyscout)
behavioral_vector	array<double>	192-dim embedding for this player-match
stat_vector	array<double>	13-dim stat vector (may be NULL)

Schema Migration — Dual-Column Window (2026-04-25 → 2026-07-22)

PR 5b of the lakehouse Kimball migration (ADR-011) adds the BIGINT surrogate player_key to the underlying fct_player_embeddings* marts. This dataset payload is NOT yet modified — the parquet files continue to ship canonical_player_id only. PR 8 (planned 2026-07-22) will add player_key to the payloads in a backwards-compatible way and announce a sunset for canonical_player_id.

Recommended consumer behaviour during this window:

• No change required. Continue to read canonical_player_id from this dataset.
• If you maintain your own join to a dim_players clone, you may pre-compute player_key = xxhash64(provider || '|' || cast(player_id as string)) to align with the lakehouse Kimball convention ahead of the payload change.
• After 2026-07-22 the dataset will carry both columns for at least one HF dataset version, then canonical_player_id will be deprecated. Migrate at your convenience inside that window.

If you depend on this dataset and need extra notice before the column drop, open an issue on the lakehouse repo.

Use Cases

• Similarity search: Find players with the most similar style to a given player using cosine similarity on behavioral_vector. Four HNSW indexes are maintained on the platform for sub-millisecond ANN queries.
• Clustering: Group players by behavioral archetype (e.g., deep-lying playmaker vs. box-to-box midfielder) without labeled data.
• Transfer market analysis: Identify stylistically equivalent players across leagues and data sources for scouting.
• Style evolution tracking: Use the season config to monitor how a player's behavioral profile changes across seasons or after a transfer.
• Hybrid ranking: Combine behavioral_vector (style match) with stat_vector (volume/efficiency) for multi-objective player ranking.

Data Sources

Source	Matches	License
StatsBomb Open Data	~3,000	CC-BY 4.0
Wyscout Public Dataset	~1,900	CC-BY-NC 4.0

Coverage includes the Premier League, La Liga, Serie A, Bundesliga, Ligue 1, Champions League, World Cup, and more.

Limitations

• Event-based only: The behavioral vector is derived from action sequences, not tracking data. Off-ball movement, pressing intensity, and spatial positioning are not directly encoded.
• Action sequence style: The transformer model captures the type and order of actions a player takes, not physical attributes (pace, strength) or tactical context (formation, instructions).
• Open data only: Model trained on publicly available StatsBomb and Wyscout data. Players with few appearances in these datasets may have noisy embeddings.
• NULL stat vectors: Players with insufficient on-ball volume for reliable per-90 computation have stat_vector = NULL. Downstream models should handle this case explicitly.
• Cross-source alignment: Player identity is unified via the entity resolution pipeline (dim_players), but subtle cross-source differences in event definitions may affect behavioral vector comparability.

Citation

If you use these embeddings, please reference the companion model repository:

luxury-lakehouse/football2vec-statsbomb-wyscout
https://huggingface.co/luxury-lakehouse/football2vec-statsbomb-wyscout

And the underlying SPADL/VAEP framework:

@inproceedings{decroos2019actions,
  title={Actions Speak Louder than Goals: Valuing Player Actions in Soccer},
  author={Decroos, Tom and Bransen, Lotte and Van Haaren, Jan and Davis, Jesse},
  booktitle={Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining},
  pages={1851--1861},
  year={2019},
  publisher={ACM}
}

More Information

Explore interactively: Soccer Analytics App

• Model repo: luxury-lakehouse/football2vec-statsbomb-wyscout
• License: MIT

PR 7 changelog (2026-04-27)

PR 5b (2026-04-25) added Kimball surrogate FK player_key to the upstream gold mart. The HF dataset payload republish was deferred at PR 5b and is absorbed into PR 7's scope per feedback_hf_artifacts_in_scope_pr and project_kimball_pr8_scope_locked. Payload now carries player_key (BIGINT) alongside the legacy player_id and canonical_player_id columns during the 2026-07-22 dual-column window. PR 8 will sunset the legacy ID columns post-2026-07-22.