luxury-lakehouse/football2vec-360-embeddings

Football2Vec 360 Player Embeddings — 208-Dim Transformer + Deep Sets Vectors

Pre-computed 208-dimensional player embedding vectors from the Football2Vec 360-Enriched model — ready to use without loading model weights. Covers ~4K player-match records across 323 StatsBomb 360 open-data matches. Each vector encodes both action sequences and spatial freeze-frame context via a transformer encoder (192d) combined with a Deep Sets encoder (16d).

This dataset occupies a separate embedding space from Football2Vec v2 Player Embeddings (192-dim). Vectors from the two models are not directly comparable and must not be mixed in the same similarity index.

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-360-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]} player-matches, {vectors.shape[1]}-dim embeddings")  # (~4K, 208)

# Cosine similarity between two players
from sklearn.metrics.pairwise import cosine_similarity
player_a = vectors[0:1]
player_b = vectors[1:2]
sim = cosine_similarity(player_a, player_b)[0, 0]
print(f"Cosine similarity: {sim:.4f}")

Explore interactively: Soccer Analytics App

What Are These Embeddings?

Each embedding is a 208-dimensional vector combining two complementary representations:

• Transformer stream (192 dimensions): A transformer encoder embedding capturing action sequences and spatial patterns from SPADL-tokenized events. Same architecture as Football2Vec v2 but trained exclusively on 360-annotated matches.
• Deep Sets stream (16 dimensions): A permutation-invariant encoder (Zaheer et al. 2017) processing the unordered set of visible player positions (freeze-frame) at each action, aggregated via sum-pooling. Captures how a player behaves relative to surrounding opponents and teammates.

Both streams are combined via concatenation and jointly trained with adversarial team debiasing (Ganin et al. 2016) to remove team-identity confounds.

For model architecture details and training methodology, see the companion model: luxury-lakehouse/football2vec-360.

Data Fields

Column	Type	Description
canonical_player_id	string	Unified player identifier (from entity resolution across data sources)
match_id	string	Match identifier (StatsBomb 360 match)
behavioral_vector	array<double>	208-dim embedding for this player-match [192d transformer || 16d Deep Sets]

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 (which produce this 360 dataset alongside the v2 dataset). 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.

Coverage

Metric	Value
Matches	323 (complete StatsBomb 360 open-data release)
Player-match records	~4K
Competitions	La Liga, Premier League, Champions League, Euro 2020, Women's World Cup, Copa America

Coverage is limited to players with appearances in StatsBomb 360-annotated matches. For broader player coverage (~87K player-matches, ~3,000 matches), use Football2Vec v2 Player Embeddings.

Use Cases

• Context-aware similarity search: Cosine distance on 208-dim vectors finds players with similar style and spatial decision-making in 360-annotated matches
• Spatial pattern analysis: The 16-dim Deep Sets component enables queries such as "players who behave similarly under high defensive pressure"
• Ablation research: Compare with Football2Vec v2 embeddings to quantify the impact of 360 freeze-frame context on player representations
• Transfer scouting: Identify players with equivalent behavioral profiles in competitions with 360 data coverage
• Downstream features: Input to GNN tactical models where spatial context and relational reasoning matter

Limitations

• 360-match coverage only: Players without StatsBomb 360 match appearances have no embeddings in this dataset. Use Football2Vec v2 embeddings for broader coverage.
• Per-match granularity: One row per player-match (no career or season aggregates in this release). Aggregate across matches client-side if needed.
• Separate embedding space: 208-dim vectors are not comparable to Football2Vec v2 192-dim vectors. Cannot mix in the same similarity index without re-embedding all players.
• Small corpus effects: 323 matches is a smaller training corpus than Football2Vec v2 (~3,000 matches). Players with few 360 appearances may have noisier embeddings.
• Open data only: Derived from publicly available StatsBomb 360 data. Commercial datasets with proprietary 360 annotations may yield different representations.

Freshness

Metric	Value
Freshness SLA	168 hours (7 days)
Inference schedule	Daily 06:00 UTC
Skip guard	match_id-level — only new 360 matches trigger re-inference

Citation

If you use these embeddings, please cite the companion model and the Deep Sets architecture:

@inproceedings{zaheer2017deep,
  title={Deep Sets},
  author={Zaheer, Manzil and Kottur, Satwik and Ravanbakhsh, Siamak and Poczos, Barnabas and Salakhutdinov, Ruslan and Smola, Alexander},
  booktitle={Advances in Neural Information Processing Systems},
  volume={30},
  year={2017}
}

@software{nielsen2026football2vec_360,
  title={Football2Vec 360-Enriched: Transformer + Deep Sets Player Embeddings},
  author={Nielsen, Karsten Skyt},
  year={2026},
  url={https://github.com/karsten-s-nielsen/luxury-lakehouse}
}

Companion Resources

Resource	Description
Football2Vec 360 Model	208-dim model that generated these embeddings
360 Training Data	SPADL sequences with freeze-frames used for training
Football2Vec v2 Embeddings	192-dim event-only embeddings with broader coverage
Football2Vec v2 Model	192-dim event-only transformer model
SPADL/VAEP Action Values	Per-action offensive/defensive VAEP valuations

More Information

Explore interactively: Soccer Analytics App

• Model repo: luxury-lakehouse/football2vec-360
• License: CC-BY 4.0 (StatsBomb Open Data)

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.