Football2Vec 360-Enriched — Transformer + Deep Sets Player Embeddings

208-dimensional player embedding vectors from a 4-layer transformer encoder augmented with a Deep Sets context encoder (Zaheer et al. 2017) trained on ~2M SPADL actions with StatsBomb 360 freeze-frame data from 323 professional soccer matches. Adversarial team debiasing via gradient reversal (Ganin et al. 2016) removes team-identity confounds, producing style representations that generalize across teams.

This model occupies a separate embedding space from Football2Vec v2 (192-dim, event-only). The 360-enriched vectors are not directly comparable to v2 vectors and should not be mixed in downstream similarity search without re-indexing.

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

Model Description

Football2Vec 360-Enriched extends the v2 transformer architecture with a Deep Sets encoder that processes the spatial positions of all visible opponents and teammates at the moment of each action. The 208-dim output captures both individual action sequences and the spatial context in which those actions occur — richer representations than event-only models for players who frequently appear in 360-annotated matches.

Architecture

Component	Detail
Token embedding	23 SPADL action types → 192d lookup table
Spatial encoding	MLP(x) + MLP(y) → 192d each, summed with token embedding
Positional embedding	Learnable, max 512 tokens
Encoder	4-layer TransformerEncoder, 4 attention heads, GELU activation, 4x FFN
Pooling	Mean pooling over valid (non-padding) tokens → 192d
Deep Sets encoder	Per-player MLP on freeze-frame (x, y, team) → sum-pool → 16d
Output	Concatenation [192d transformer \|\| 16d Deep Sets] → 208d
Adversarial head	Gradient reversal layer (λ=0.2) + team classifier

Two-Stage Training

Stage 1 — Masked Language Modeling: 15% of action tokens are masked; the model predicts the original action type from surrounding context. The Deep Sets encoder processes freeze-frame coordinates at the masked token position, providing spatial context to the transformer.

Stage 2 — Adversarial Debiasing: A team classifier head is attached via a gradient reversal layer (Ganin et al. 2016). The encoder learns to produce embeddings that cannot predict which team a player belongs to, removing team-system confounds while retaining individual style signal.

Dual-Vector Architecture

This model provides the behavioral half of a dual-vector player representation:

Vector	Dimensions	Source	Captures
Behavioral (this model)	208	Transformer + Deep Sets on SPADL + 360 freeze-frames	Playing style, spatial context, action sequences
Statistical	13	Z-score normalized per-90 stats	Goals, assists, xG, passes, VAEP, defensive metrics

Both vectors are stored in PostgreSQL with pgvector HNSW indexes for sub-10ms similarity queries.

Training Data

Source	Matches	Events	License
StatsBomb 360 Open Data	323	~2M	CC-BY 4.0

The 323-match corpus is the complete StatsBomb 360 open-data release. Coverage includes La Liga, Premier League, Champions League, Euro 2020, and Women's World Cup matches with freeze-frame annotations.

Training data is published as luxury-lakehouse/football2vec-360-training-data on HF Hub.

Tokenization

Events are tokenized using the 23-type SPADL vocabulary: pass, cross, throw_in, freekick_crossed, freekick_short, corner_crossed, corner_short, take_on, foul, tackle, interception, shot, shot_penalty, shot_freekick, keeper_save, keeper_claim, keeper_punch, keeper_pick_up, clearance, bad_touch, non_action, dribble, goalkick.

Continuous spatial coordinates (x, y) normalized to [0, 1] on a 105×68m pitch are injected via learned MLP projections. Freeze-frame coordinates encode visible players as an unordered set with a team indicator bit (0 = teammate, 1 = opponent).

Hyperparameters

Parameter	Value
Hidden dimension	192
Encoder layers	4
Attention heads	4
FFN multiplier	4x (768)
Dropout	0.1
Max sequence length	512
MLM mask probability	0.15
Spatial MLP intermediate dim	64
Deep Sets MLP dims	[32, 16]
Output dimension	208
Batch size	256
Learning rate	1e-4
Weight decay	0.01
Warmup fraction	10%
Adversarial λ max	0.2
Adversarial warmup epochs	5

Training runs on HF Jobs A10G-small GPU (~90 minutes total for both stages).

How to Use

Quick Start

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import json

# Download model weights (Stage 2 — adversarial debiased)
weights_path = hf_hub_download("luxury-lakehouse/football2vec-360", "stage2/model.safetensors")
config_path = hf_hub_download("luxury-lakehouse/football2vec-360", "stage2/config.json")

with open(config_path) as f:
    config = json.load(f)

state_dict = load_file(weights_path)
print(f"Config: {config['hidden_dim']}-dim transformer + {config['deepsets_dim']}-dim Deep Sets")
print(f"Output dimension: {config['output_dim']}")  # 208
print(f"Parameters: {sum(p.numel() for p in state_dict.values()):,}")

Pre-Computed Embeddings (recommended)

For most use cases, load the pre-computed embeddings directly — no model inference needed:

from datasets import load_dataset
import numpy as np

ds = load_dataset("luxury-lakehouse/football2vec-360-embeddings")
df = ds["train"].to_pandas()

vectors = np.array(df["behavioral_vector"].tolist())
print(f"{vectors.shape[0]} player-matches, {vectors.shape[1]}-dim embeddings")  # (~4K, 208)

Note: These embeddings cover only players with StatsBomb 360 match appearances (~4K player-match records vs. ~87K for Football2Vec v2). For broader coverage, use Football2Vec v2.

Intended Use

Context-aware player similarity: Cosine distance on 208-dim vectors finds players with similar style and spatial decision-making
Spatial pattern analysis: The 16-dim Deep Sets component captures how players behave relative to nearby opponents and teammates
Scouting in high-press contexts: Embeddings encode how players handle actions under spatial pressure from surrounding defenders
Research: Reproducible 360-enriched player representations with adversarial debiasing; pairs with Football2Vec v2 for ablation studies
Downstream features: Input to GNN tactical models where spatial context matters

EU AI Act — Intended Use and Non-Use

This model is published for research and reproducibility purposes on public, open-licensed match data. It is not intended for, not validated for, and not supplied to any use that would fall within Annex III §4 (Employment, workers management and access to self-employment) of Regulation (EU) 2024/1689 — including recruitment or selection of natural persons, decisions affecting work-related contractual relationships, promotion, termination, task allocation based on individual traits, or the monitoring and evaluation of performance and behaviour of workers for employment decisions. Player similarity search is a canonical scouting workflow, and any deployer is responsible for treating this model as decision-support at most, never as a decision system.

Any deployer who wishes to use this model for such a purpose is responsible for performing their own conformity assessment under Article 43, for drawing up the technical documentation required by Article 11 and Annex IV, for implementing the human oversight measures required by Article 14, for declaring accuracy metrics under Article 15, and for ensuring the data governance obligations of Article 10 are met. Note specifically that the training data contains no protected attributes and therefore cannot support the group-fairness audits required by Article 10(2)(g) without ingesting additional personal data. The adversarial team debiasing (Ganin et al. 2016) described above addresses a confounding effect (team-system leakage) and is not a substitute for an Article 10 protected-attribute audit.

See the AI_GOVERNANCE.md gap analysis in the source repository for the project's full risk classification, re-classification triggers, and governance posture.

Limitations

360-data only: Covers 323 StatsBomb 360 matches. Players with appearances only in non-360 matches have no embeddings from this model.
Smaller training corpus: 323 matches vs. ~3,000 for Football2Vec v2. Embeddings for players with few 360 appearances may be noisier.
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.
Event-based actions + freeze-frames: Off-ball runs and pressing without a nearby action event are not captured.
Team debiasing, not competition debiasing: The adversarial head targets team ID (stronger confounder in the smaller 360 corpus). Cross-league confounds are attenuated but not fully removed.
Open data only: Derived from publicly available StatsBomb 360 data. Commercial datasets with proprietary 360 annotations may yield different representations.

Freshness

Metric	Value
Training data freshness SLA	168 hours (7 days)
Inference schedule	Daily 06:00 UTC
Skip guard	`match_id`-level — only new 360 matches are processed

Model Files

stage1/model.safetensors      -- Stage 1 MLM checkpoint (safetensors format)
stage2/model.safetensors      -- Stage 2 adversarial, final (safetensors format)
stage2/config.json            -- Football2Vec360Config as JSON
zscore_params.json            -- z-score normalization parameters (13-dim stat vector)

Model weights use the safetensors format — a tensor-only serialization with zero pickle surface and no code execution capability. Pre-computed embeddings are delivered as Parquet (non-executable).

Citation

@inproceedings{theiner2022explainable,
  title={Explainable Expected Goal Models for Performance Analysis in Football},
  author={Theiner, Jonas and M{\"u}ller-Budack, Eric and Ewerth, Ralph},
  booktitle={Proceedings of the 4th International Workshop on Multimedia Content Analysis in Sports},
  pages={39--47},
  year={2022}
}

@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}
}

@article{ganin2016domain,
  title={Domain-Adversarial Training of Neural Networks},
  author={Ganin, Yaroslav and Ustinova, Evgeniya and Cambau, Hana
          and Lempitsky, Victor and Laviolette, Fran{\c{c}}ois},
  journal={Journal of Machine Learning Research},
  volume={17},
  number={1},
  pages={1--35},
  year={2016}
}

@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
360 Training Data	SPADL sequences with 360 freeze-frames used for training
360 Player Embeddings	Pre-computed 208-dim vectors per player-match
Football2Vec v2	192-dim event-only model (~3,000 matches, broader coverage)
Football2Vec v1	32-dim Doc2Vec baseline
SPADL/VAEP Action Values	Per-action offensive/defensive VAEP valuations

Demo

Try the interactive Soccer Analytics App — the Player Similarity page supports similarity search on 360-enriched embeddings for players with 360 match coverage.

Explore interactively: Soccer Analytics App

More Information

License: CC-BY-NC 4.0 (inherited from StatsBomb open data terms)
v2 event-only model: Football2Vec v2
v1 baseline model: Football2Vec v1 (Doc2Vec)
Platform: Luxury Lakehouse Soccer Analytics
Workflow card: workflow-cards/wf-football2vec-360.yaml

Downloads last month: -; Downloads are not tracked for this model. How to track

luxury-lakehouse
/

football2vec-360