spadl-action-context
收藏资源简介:
SPADL Action Context Features数据集是一个来自luxury-lakehouse分析平台的足球比赛统一每动作上下文特征数据集。该数据集专门为包含帧数据的比赛设计(遵循frames-required原则,ADR-057),每行对应一个SPADL(Soccer Player Action Description Language)动作。数据集包含两个层级的数据提供商:追踪数据提供商(IDSSE、Metrica、SkillCorner、GradientSports)提供完整的追踪列特征,包括控球、团队形态、线路突破、OBSO、PAUSA、守门员影响力区域、xShotOccurrence、结构化上下文、xCross、球员影响力、射门球门交叉几何、守门员传球完成率(`gk_completion_*`)以及xT-GK守门员分布系列(`xt_gk_*`,包含五种预设哲学);StatsBomb 360冻结帧数据(SB360)提供部分/稀疏的单帧追踪列子集,由于缺乏速度数据,控球模型使用Voronoi方法,速度依赖的特征系列(DAS、覆盖阴影、离球、xCross、射门前守门员位置)为空值。数据集通过`pitch_control_method`列记录控球模型来源(追踪数据使用spearman,SB360使用voronoi),通过`ghost_gk_method`列记录幽灵守门员KDE后端来源。数据集规模为10万到100万行之间,包含多个配置:all(默认,所有公开提供商)、idsse、metrica、skillcorner。GradientSports分区因许可证限制位于私有伴生仓库中。数据字段包括身份列(数据源、比赛ID、动作ID、周期ID、团队ID、球员ID等字符串和整型)、游戏状态列(时间、动作类型、起始和结束坐标等)、帧链接列以及大量上下文特征列(全部为float64类型,可为空NaN)。该数据集适用于足球分析任务,特别是需要结合动作和上下文特征的高级分析,如控球分析、空间创造评估、守门员影响力建模、射门发生概率预测等。
The SPADL Action Context Features dataset is a unified per-action context feature dataset for soccer matches from the luxury-lakehouse analytics platform. It is specifically designed for matches with frame data (following the frames-required principle, ADR-057), with each row corresponding to a SPADL (Soccer Player Action Description Language) action. The dataset includes two levels of data providers: tracking data providers (IDSSE, Metrica, SkillCorner, GradientSports) provide complete tracking column features, including pitch control, team formation, line breaking, OBSO, PAUSA, goalkeeper influence zones, xShotOccurrence, structured context, xCross, player influence, shot-goal intersection geometry, goalkeeper pass completion rates (`gk_completion_*`), and xT-GK goalkeeper distribution series (`xt_gk_*`, with five preset philosophies); StatsBomb 360 freeze-frame data (SB360) provides a partial/sparse subset of single-frame tracking columns, with speed-dependent feature series (DAS, coverage shadows, off-ball, xCross, pre-shot goalkeeper positioning) being null due to lack of speed data, and pitch control model using the Voronoi method. The dataset records the pitch control model source via the `pitch_control_method` column (spearman for tracking data, voronoi for SB360) and the ghost goalkeeper KDE backend source via the `ghost_gk_method` column. The dataset size ranges from 100,000 to 1,000,000 rows, with multiple configurations: all (default, all public providers), idsse, metrica, skillcorner. The GradientSports partition is in a private companion repository due to licensing restrictions. Data fields include identity columns (data source, match ID, action ID, period ID, team ID, player ID, etc., as strings and integers), game state columns (time, action type, start and end coordinates, etc.), frame linking columns, and numerous context feature columns (all float64 type, nullable as NaN). The dataset is suitable for soccer analytics tasks, especially advanced analysis requiring combined action and context features, such as pitch control analysis, space creation evaluation, goalkeeper influence modeling, and shot occurrence probability prediction.
数据集概要
数据集名称:SPADL Action Context Features
数据集地址:https://huggingface.co/datasets/luxury-lakehouse/spadl-action-context
许可证:CC-BY-NC-4.0
数据集规模:100K < n < 1M
标签:soccer, football, analytics, spadl, action-context, tracking, pitch-control, expected-threat
数据集描述
该数据集基于足球分析平台 luxury-lakehouse 构建,为每一条 SPADL 动作提供统一的环境特征。每个动作对应一行数据。数据集仅包含具有追踪数据(Tracking)或 StatsBomb 360 冻结帧(SB360)的比赛。仅包含事件数据的比赛(如 Wyscout、无 360 的 StatsBomb)不在此数据集中。
数据提供商层级
| 层级 | 提供商 | 特征覆盖范围 |
|---|---|---|
| 追踪(Tracking) | IDSSE, Metrica, SkillCorner, GradientSports | 完整追踪特征:球场控制、球队阵型、突破线、OBSO、PAUSA、守门员影响区域、xShotOccurrence、结构上下文、xCross、球员影响、射门球门线穿越几何、守门员传球完成率(gk_completion_)、以及 xT-GK 守门员分布家族(xt_gk_,五种哲学预设),包括四个解析坐标审计列(xt_gk_origin_x/y + xt_gk_dest_x/y) |
| SB360 | StatsBomb(含360冻结帧) | 部分/稀疏单帧追踪特征子集——冻结帧无速度数据,因此球场控制为 voronoi,速度依赖型特征(DAS、cover-shadows、off-ball、xCross、射门前守门员位置)为 null |
关键方法说明
- 球场控制方法(pitch_control_method):追踪提供商使用 speedman(速度感知模型),SB360 使用 voronoi(仅位置模型),两者不可直接比较,需按此列分组。
- 幽灵守门员方法(ghost_gk_method):记录生成 ghost_gk_* 列的 KDE 后端(如 scipy, fft-cic 等),不同后端值不同,需按此列分组后再比较。
- SB360 冻结帧覆盖:部分且稀疏,每条指标仅对包含所需球员的冻结帧填充,xshot_occurrence 仅对约 4% 的 SB360 动作非空(非随机子样本)。计算 SB360 指标时需过滤非空值并考虑采样偏差。
数据集结构与加载
数据集按提供商分为多个配置(config),可单独加载:
python from datasets import load_dataset
加载所有公开提供商(config "all")
ds = load_dataset("luxury-lakehouse/spadl-action-context") df = ds["train"].to_pandas()
仅加载 SkillCorner
sc = load_dataset("luxury-lakehouse/spadl-action-context", "skillcorner")["train"].to_pandas()
可用配置:all(默认)、idsse、metrica、skillcorner。GradientSports 由于许可证限制,存储于私有仓库 spadl-action-context-restricted。
数据列分类
| 类别 | 示例列 |
|---|---|
| 身份 | data_source, match_id, action_id, period_id, team_id, player_id |
| 比赛状态 | time_seconds, type_name, start_x, start_y, end_x, end_y |
| 帧关联 | frame_id, timestamp_utc |
| 守门员解析 | defending_gk_player_id_native |
| 守门员空间 | gk_x, gk_y, gk_distance_to_goal |
| 动作上下文 | nearest_defender_distance, receiver_zone_density, defenders_in_triangle_to_goal |
| 动作前窗口 | actor_arc_length_pre_window, actor_displacement_pre_window |
| 压力 | pressure_on_actor__andrienko_oval, __link_zones, __bekkers_pi |
| 球场控制 | pitch_control_at_target__spearman, __fernandez_bornn, __voronoi |
| 防线 | defensive_line_x, back_line_high_x, compactness_x, lateral_width |
| 无球上下文 | line_break, n_attackers_behind_line, n_off_ball_runners_* |
| Ward 突破线 | line_break__ward, lines_broken__ward, line_breaking_type__ward |
| 球队阵型 | team_shape_{metric}_{attacking/defending}(14列) |
| DAS | das_team, das_opponent, das_diff |
| 守门员影响 | gk_pitch_control_share_weighted, gk_reachable_area_m2, gk_closing_time_* |
| 覆盖阴影 | n_blocked_receivers, blocking_score, blocked_threat_fraction |
| 同步分数 | sync_score_min, sync_score_mean, sync_score_high_quality_frac |
| OBSO | obso_value, obso_total_threat |
| PAUSA | pausa_value, pausa_added_threat |
| 空间创造 | space_created_m2, space_denied_m2_opponent |
| ELASTIC 同步 | elastic_sync_score, elastic_compactness |
| 形状图 | shape_graph_centrality, shape_graph_clustering |
| 幽灵守门员 | ghost_gk_x, ghost_gk_y, ghost_gk_density_spread |
| 结构传球 | structural_lbs, structural_sgm, structural_sdi |
| 球员影响 | actor_reachable_area_m2, off_ball_xt_{team,opponent,diff}, reachable_area_{team,opponent,diff} |
| xCrossAttempt | xcross_attempt |
| xShotOccurrence | xshot_occurrence |
| 射门球门线 | shot_crossing_y, shot_crossing_z, shot_speed, shot_time_to_goal_line, shot_on_target_derived, shot_z_profile |
| 球场控制方法原点 | pitch_control_method |
| 幽灵守门员方法原点 | ghost_gk_method |
数据字段类型
所有特征列均为 float64(可空 NaN),身份列类型各异:data_source(字符串)、match_id(字符串)、action_id(整数)、period_id(整数)、time_seconds(浮点数)、team_id(字符串)、player_id(字符串)、type_name(字符串)、start_x/start_y/end_x/end_y(浮点数,SPADL 105x68 坐标系)。
引用
bibtex @software{luxury_lakehouse, title = {Luxury Lakehouse — Serverless Soccer Analytics Platform}, url = {https://github.com/karsten-s-nielsen/luxury-lakehouse} }
许可证
CC-BY-NC-4.0




