soccer-feature-engineering / soccer_feature_engineering.py
knoxel's picture
Upload soccer_feature_engineering.py
56d87f4 verified
"""
Soccer Feature Engineering Pipeline - Competition Version
===========================================================
Engineers 33 match-level, team-level features from SkillCorner dynamic_events.csv.
Based on the Kaggle Soccer Feature Engineering Hackathon requirements.
This script uses the EXACT column naming convention from the reference notebook
by Dev0907 to ensure full compatibility with competition evaluation.
Input: SkillCorner opendata repository (or any folder with *_dynamic_events.csv)
Output: features.csv - one row per team per match
"""
import glob
import os
import pandas as pd
def discover_dynamic_events_files(data_root="/app/opendata/data/matches"):
"""Dynamically discover all *_dynamic_events.csv files via glob."""
pattern = os.path.join(data_root, "**", "*_dynamic_events.csv")
files = glob.glob(pattern, recursive=True)
files.sort()
return files
def compute_team_features(df, team):
"""
Compute 33 aggregated match-level features for a single team in a single match.
Uses EXACT column names from the reference notebook by Dev0907.
"""
match_id = df["match_id"].iloc[0]
team_df = df[df["team_id"] == team]
poss = team_df[team_df["event_type"] == "player_possession"]
obe = team_df[team_df["event_type"] == "on_ball_engagement"]
obr = team_df[team_df["event_type"] == "off_ball_run"]
passes = poss[poss["pass_outcome"].notna()]
# DIMENSION 1: ATTACKING STRUCTURE
att1 = int((passes["third_end"] == "attacking_third").sum())
carries = poss[poss["carry"] == True]
att2 = int((carries["third_end"] == "attacking_third").sum())
att3 = int(passes["n_opponents_bypassed"].clip(lower=0).sum())
att4 = int(passes["last_line_break"].sum()) if "last_line_break" in passes.columns else 0
att5 = int((passes["third_start"] == "attacking_third").sum())
# DIMENSION 2: BUILD-UP PROFILE
att6 = int((poss["team_in_possession_phase_type"] == "build_up").sum())
att7 = int((poss["team_in_possession_phase_type"] == "direct").sum())
att8 = int((poss["team_in_possession_phase_type"] == "set_play").sum())
att9 = int((poss["team_in_possession_phase_type"] == "quick_break").sum())
att10 = int((poss["team_in_possession_phase_type"] == "transition").sum())
# DIMENSION 3: POSSESSION QUALITY
att11 = int(poss["one_touch"].sum())
att12 = int(poss["quick_pass"].sum())
att13 = int(poss["lead_to_shot"].sum())
att14 = int(poss["lead_to_goal"].sum())
att15 = float(round(poss["delta_to_last_defensive_line_gain"].clip(lower=0).sum(), 2))
att16 = float(round(poss["last_defensive_line_height_gain"].clip(lower=0).sum(), 2))
att17 = int(poss["forward_momentum"].sum())
att18 = int(poss["n_passing_options"].sum())
att19 = int(poss["n_passing_options_dangerous_difficult"].sum())
att20 = int((obr["event_subtype"] == "run_ahead_of_the_ball").sum())
# DIMENSION 4: PRESSING & DEFENSIVE SHAPE
def1 = int(len(obe))
def2 = int((obe["event_subtype"] == "counter_press").sum())
def3 = int((obe["event_subtype"] == "recovery_press").sum())
chains = obe[obe["pressing_chain_length"].notna()]
chain_starts = chains[chains["index_in_pressing_chain"] == 1.0]
def4 = int(chain_starts["pressing_chain_length"].sum())
def5 = int(len(chain_starts))
def6 = int(chains["pressing_chain_length"].max()) if len(chains) > 0 else 0
def7 = int(obe["stop_possession_danger"].sum())
# DIMENSION 5: OFF-BALL MOVEMENT INTELLIGENCE
run1 = int(obr["break_defensive_line"].sum())
run2 = int(obr["push_defensive_line"].sum())
run3 = int((obr["event_subtype"] == "behind").sum())
run4 = int((obr["event_subtype"] == "overlap").sum())
run5 = int((obr["third_start"] == "attacking_third").sum())
gng = int(poss["initiate_give_and_go"].sum()) if "initiate_give_and_go" in poss.columns else 0
return {
"match_id": match_id,
"team_id": team,
# Attacking Structure
"att1_passes_into_final_third": att1,
"att2_carries_into_attacking_third": att2,
"att3_opponents_bypassed_by_passes": att3,
"att4_last_line_break_passes": att4,
"att5_passes_in_attacking_third": att5,
# Build-Up Profile
"att6_buildup_phase_events": att6,
"att7_direct_phase_events": att7,
"att8_setplay_events": att8,
"att9_quickbreak_events": att9,
"att10_transition_events": att10,
# Possession Quality
"att11_one_touch_passes": att11,
"att12_quick_passes": att12,
"att13_possessions_leading_to_shot": att13,
"att14_possessions_leading_to_goal": att14,
"att15_def_line_depth_total_pushed_m": att15,
"att16_def_line_height_total_pushed_m": att16,
"att17_forward_momentum_possessions": att17,
"att18_passing_options_total": att18,
"att19_dangerous_difficult_pass_options": att19,
"att20_runs_ahead_of_ball": att20,
# Defensive Pressing
"def1_total_defensive_engagements": def1,
"def2_counter_press_actions": def2,
"def3_recovery_press_actions": def3,
"def4_pressing_chain_total_length": def4,
"def5_pressing_chains_initiated": def5,
"def6_max_pressing_chain_length": def6,
"def7_danger_stopped": def7,
# Off-Ball Movement
"run1_line_breaking_runs": run1,
"run2_line_pushing_runs": run2,
"run3_runs_behind_defense": run3,
"run4_overlap_runs": run4,
"run5_attacking_third_runs": run5,
"att_give_and_go_initiated": gng,
}
def run_pipeline(data_root="/app/opendata/data/matches", output_path="/app/features.csv"):
files = discover_dynamic_events_files(data_root)
print(f"Discovered {len(files)} dynamic_events.csv files")
records = []
for f in files:
try:
df = pd.read_csv(f, low_memory=False)
match_id = df["match_id"].iloc[0]
teams = sorted(df["team_id"].unique().tolist())
for team in teams:
features = compute_team_features(df, team)
records.append(features)
print(f" match {match_id}: {len(teams)} teams")
except Exception as e:
print(f" ERROR {os.path.basename(f)}: {e}")
if not records:
raise ValueError("No valid match files processed.")
features_df = pd.DataFrame(records)
# Ensure exact column order as in reference notebook
col_order = [
"match_id", "team_id",
"att1_passes_into_final_third",
"att2_carries_into_attacking_third",
"att3_opponents_bypassed_by_passes",
"att4_last_line_break_passes",
"att5_passes_in_attacking_third",
"att6_buildup_phase_events",
"att7_direct_phase_events",
"att8_setplay_events",
"att9_quickbreak_events",
"att10_transition_events",
"att11_one_touch_passes",
"att12_quick_passes",
"att13_possessions_leading_to_shot",
"att14_possessions_leading_to_goal",
"att15_def_line_depth_total_pushed_m",
"att16_def_line_height_total_pushed_m",
"att17_forward_momentum_possessions",
"att18_passing_options_total",
"att19_dangerous_difficult_pass_options",
"att20_runs_ahead_of_ball",
"def1_total_defensive_engagements",
"def2_counter_press_actions",
"def3_recovery_press_actions",
"def4_pressing_chain_total_length",
"def5_pressing_chains_initiated",
"def6_max_pressing_chain_length",
"def7_danger_stopped",
"run1_line_breaking_runs",
"run2_line_pushing_runs",
"run3_runs_behind_defense",
"run4_overlap_runs",
"run5_attacking_third_runs",
"att_give_and_go_initiated",
]
features_df = features_df[col_order]
features_df.to_csv(output_path, index=False)
print(f"\nWrote {len(features_df)} rows x {len(features_df.columns)} columns to {output_path}")
print(f"Shape: {features_df.shape}")
return features_df
if __name__ == "__main__":
run_pipeline()