luxury-lakehouse/football2vec-360-embeddings
收藏资源简介：
--- language: [en] license: cc-by-4.0 task_categories: [feature-extraction] tags: [sports-analytics, soccer, football, player-embeddings, transformer, deep-sets, 360-data, similarity-search] size_categories: [1K-10K] configs: - config_name: default data_files: - split: train path: "data/*.parquet" default: true --- # Football2Vec 360 Player Embeddings — 144-Dim Transformer + Deep Sets Vectors Pre-computed 144-dimensional player embedding vectors from the [Football2Vec 360-Enriched](https://huggingface.co/luxury-lakehouse/football2vec-360) 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 (128d) combined with a Deep Sets encoder (16d). This dataset occupies a **separate embedding space** from [Football2Vec v2 Player Embeddings](https://huggingface.co/datasets/luxury-lakehouse/football2vec-player-embeddings) (128-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 ```python 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, 144) # 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:** [HF Space demo](https://huggingface.co/spaces/luxury-lakehouse/soccer-analytics-demo) ## What Are These Embeddings? Each embedding is a 144-dimensional vector combining two complementary representations: - **Transformer stream** (128 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`](https://huggingface.co/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>` | 144-dim embedding for this player-match [128d transformer \|\| 16d Deep Sets] | ## 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](https://huggingface.co/datasets/luxury-lakehouse/football2vec-player-embeddings). ## Use Cases - **Context-aware similarity search**: Cosine distance on 144-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**: 144-dim vectors are not comparable to Football2Vec v2 128-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: ```bibtex @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} } ``` ```bibtex @software{nielsen2026football2vec_360, title={Football2Vec 360-Enriched: Transformer + Deep Sets Player Embeddings}, author={Nielsen, Karsten Skytt}, year={2026}, url={https://github.com/karsten-s-nielsen/luxury-lakehouse} } ``` ## Companion Resources | Resource | Description | |----------|-------------| | [Football2Vec 360 Model](https://huggingface.co/luxury-lakehouse/football2vec-360) | 144-dim model that generated these embeddings | | [360 Training Data](https://huggingface.co/datasets/luxury-lakehouse/football2vec-360-training-data) | SPADL sequences with freeze-frames used for training | | [Football2Vec v2 Embeddings](https://huggingface.co/datasets/luxury-lakehouse/football2vec-player-embeddings) | 128-dim event-only embeddings with broader coverage | | [Football2Vec v2 Model](https://huggingface.co/luxury-lakehouse/football2vec-v2) | 128-dim event-only transformer model | | [SPADL/VAEP Action Values](https://huggingface.co/datasets/luxury-lakehouse/spadl-vaep-action-values) | Per-action offensive/defensive VAEP valuations | ## More Information > **Explore interactively:** [HF Space demo](https://huggingface.co/spaces/luxury-lakehouse/soccer-analytics-demo) - **Model repo**: [`luxury-lakehouse/football2vec-360`](https://huggingface.co/luxury-lakehouse/football2vec-360) - **License**: [CC-BY 4.0](https://creativecommons.org/licenses/by/4.0/) (StatsBomb Open Data)
--- language: [英语] license: CC-BY-4.0 task_categories: [特征提取] tags: [体育分析、足球、球员嵌入、Transformer、Deep Sets、360度数据、相似度搜索] size_categories: [1000-10000] configs: - config_name: default data_files: - split: 训练集 path: "data/*.parquet" default: true --- # Football2Vec 360 球员嵌入 —— 144维Transformer与Deep Sets向量 本数据集包含从[Football2Vec 360-Enriched](https://huggingface.co/luxury-lakehouse/football2vec-360)模型预计算得到的144维球员嵌入向量，无需加载模型权重即可直接使用。数据集覆盖323场StatsBomb 360公开数据赛事中的**约4000条球员-赛事记录**。每条向量通过结合Transformer编码器（128维）与Deep Sets编码器（16维），同时编码动作序列与空间定格帧上下文信息。 本数据集的嵌入空间与[Football2Vec v2 球员嵌入](https://huggingface.co/datasets/luxury-lakehouse/football2vec-player-embeddings)（128维）相互独立。两种模型生成的向量无法直接比较，且不得混入同一相似度索引中。 本数据集隶属于（没错，主打轻奢！）Lakehouse足球分析平台。 ## 快速入门 python 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, 144) # 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}") > **交互式探索：** [HF Space演示站](https://huggingface.co/spaces/luxury-lakehouse/soccer-analytics-demo) ## 嵌入向量说明 每条嵌入向量均为144维，由两种互补的表征拼接而成： - **Transformer分支（128维）**：Transformer编码器生成的嵌入，从SPADL标记的赛事事件中捕捉动作序列与空间模式。其架构与Football2Vec v2一致，但仅在360度标注赛事上进行训练。 - **Deep Sets分支（16维）**：一种置换不变编码器（Zaheer等人，2017），用于处理每次动作时刻可见球员位置的无序集合（定格帧），并通过求和池化进行聚合。该分支可捕捉球员相对于周边对手与队友的行为模式。 两个分支通过拼接合并，并通过对抗性球队去偏训练（Ganin等人，2016）联合优化，以消除球队身份带来的混淆因素。 如需了解模型架构细节与训练方法，请参阅配套模型：[`luxury-lakehouse/football2vec-360`](https://huggingface.co/luxury-lakehouse/football2vec-360)。 ## 数据字段 | 列名 | 数据类型 | 描述 | |--------|------|-------------| | `canonical_player_id` | `string` | 统一球员标识符（来自多数据源实体匹配） | | `match_id` | `string` | 赛事标识符（对应StatsBomb 360赛事） | | `behavioral_vector` | `array<double>` | 该球员-赛事记录的144维嵌入向量 [128维Transformer || 16维Deep Sets] | ## 覆盖范围 | 指标 | 数值 | |--------|-------| | **赛事场数** | 323场（完整StatsBomb 360公开数据集） | | **球员-赛事记录数** | 约4000条 | | **赛事联赛** | 西甲、英超、欧冠、2020欧洲杯、女足世界杯、美洲杯 | 本数据集仅覆盖在StatsBomb 360标注赛事中登场的球员。如需更广泛的球员覆盖（约87000条球员-赛事记录、约3000场赛事），请使用[Football2Vec v2 球员嵌入](https://huggingface.co/datasets/luxury-lakehouse/football2vec-player-embeddings)。 ## 应用场景 - **上下文感知相似度搜索**：基于144维向量的余弦距离可找到在360度标注赛事中具有相似比赛风格与空间决策模式的球员 - **空间模式分析**：16维Deep Sets分支支持诸如“在高压防守环境下行为模式相似的球员”这类查询 - **消融实验研究**：与Football2Vec v2嵌入向量对比，可量化360度定格帧上下文对球员表征的影响 - **转会球探**：在覆盖360度数据的赛事中，识别行为特征匹配的球员 - **下游特征输入**：作为空间上下文与关系推理所需的GNN战术模型的输入 ## 局限性 - **仅覆盖360度标注赛事**：未在StatsBomb 360赛事中登场的球员无对应嵌入向量。如需更广泛覆盖，请使用Football2Vec v2嵌入向量。 - **单赛事颗粒度**：每条记录对应一名球员单场赛事的数据（本版本未提供生涯或赛季聚合数据）。如需聚合数据，请在客户端自行跨赛事汇总。 - **独立嵌入空间**：144维向量与Football2Vec v2的128维向量无法直接比较，若需在同一相似度索引中使用，需先对所有球员向量进行重新嵌入。 - **小样本语料效应**：323场赛事的训练语料规模小于Football2Vec v2（约3000场赛事）。仅在少量360度赛事登场的球员，其嵌入向量可能存在较高噪声。 - **仅公开数据来源**：本数据集基于公开可用的StatsBomb 360数据生成。使用带有专有360度标注的商业数据集可能得到不同的表征结果。 ## 更新时效性 | 指标 | 数值 | |--------|-------| | **更新服务等级协议** | 168小时（7天） | | **推理调度计划** | 每日UTC 06:00 | | **更新触发条件** | 以`match_id`为粒度，仅当有新增360度赛事时才会重新执行推理 | ## 引用声明 若您使用本数据集的嵌入向量，请引用配套模型与Deep Sets架构相关论文： bibtex @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} } bibtex @software{nielsen2026football2vec_360, title={Football2Vec 360-Enriched: Transformer + Deep Sets Player Embeddings}, author={Nielsen, Karsten Skytt}, year={2026}, url={https://github.com/karsten-s-nielsen/luxury-lakehouse} } ## 配套资源 | 资源 | 描述 | |----------|-------------| | [Football2Vec 360 模型](https://huggingface.co/luxury-lakehouse/football2vec-360) | 生成本数据集嵌入向量的144维模型 | | [360度训练数据](https://huggingface.co/datasets/luxury-lakehouse/football2vec-360-training-data) | 用于训练的带有定格帧的SPADL序列数据 | | [Football2Vec v2 嵌入向量](https://huggingface.co/datasets/luxury-lakehouse/football2vec-player-embeddings) | 覆盖范围更广的128维仅赛事事件嵌入向量 | | [Football2Vec v2 模型](https://huggingface.co/luxury-lakehouse/football2vec-v2) | 128维仅赛事事件Transformer模型 | | [SPADL/VAEP 动作价值数据](https://huggingface.co/datasets/luxury-lakehouse/spadl-vaep-action-values) | 单动作的进攻/防守VAEP估值数据 | ## 更多信息 > **交互式探索：** [HF Space演示站](https://huggingface.co/spaces/luxury-lakehouse/soccer-analytics-demo) - **模型仓库**：[`luxury-lakehouse/football2vec-360`](https://huggingface.co/luxury-lakehouse/football2vec-360) - **许可证**：[CC-BY 4.0](https://creativecommons.org/licenses/by/4.0/)（基于StatsBomb公开数据）
