mlp_with_kpis_xg

MLP xG Prediction Model (with Player KPIs)

This is a Multi-Layer Perceptron (MLP) model trained to predict Expected Goals (xG) in football/soccer. This version includes player-level KPIs (shots_per_90 and goals_per_90) as additional features to capture individual player shooting ability.

Model Description

• Architecture: Multi-Layer Perceptron with 3 hidden layers
• Hidden Dimensions: 128 → 64 → 32
• Input Features: 25 features (including player KPIs)
• Output: Binary probability (goal vs no goal)
• Framework: PyTorch
• Dropout Rate: 0.3
• Activation: ReLU (hidden layers), Sigmoid (output)
• Normalization: Batch Normalization after each hidden layer

Key Enhancement

This model extends the baseline MLP by incorporating player-specific KPIs:

• shots_per_90: Average shots per 90 minutes played by the player
• goals_per_90: Average goals per 90 minutes played by the player
• goals_percentage: Percentage of shots that result in goals

These features help the model capture individual player shooting quality and finishing ability beyond just the shot characteristics.

Performance Metrics

• Accuracy: 0.8839
• Precision: 0.6456
• Recall: 0.1085
• F1 Score: 0.1858
• ROC AUC: 0.7978
• Log Loss: 0.3023

Features

The model uses the following 25 features:

Shot Context Features:

• angle_to_gk
• angle_to_goal
• ball_closer_than_gk
• body_part_name_Left Foot
• body_part_name_Other
• body_part_name_Right Foot
• dist_to_gk
• distance_to_goal
• goal_dist_to_gk
• minute
• nearest_opponent_dist
• nearest_teammate_dist
• opponents_within_5m
• play_pattern_name_From Counter
• play_pattern_name_From Free Kick
• play_pattern_name_From Goal Kick
• play_pattern_name_From Keeper
• play_pattern_name_From Kick Off
• play_pattern_name_From Throw In
• play_pattern_name_Other
• play_pattern_name_Regular Play
• teammates_within_5m
• goals_percentage

Player KPI Features:

• shots_per_90
• goals_per_90

Usage

import torch
import joblib
import pandas as pd
from huggingface_hub import hf_hub_download

# Download files
model_path = hf_hub_download(repo_id="rokati/mlp_with_kpis_xg", filename="best_mlp_model.pth")
architecture_path = hf_hub_download(repo_id="rokati/mlp_with_kpis_xg", filename="model_architecture.py")
scaler_path = hf_hub_download(repo_id="rokati/mlp_with_kpis_xg", filename="scaler.pkl")
config_path = hf_hub_download(repo_id="rokati/mlp_with_kpis_xg", filename="config.json")

# Load architecture
import importlib.util
spec = importlib.util.spec_from_file_location("model_architecture", architecture_path)
model_module = importlib.util.module_from_spec(spec)
spec.loader.exec_module(model_module)

# Load model
model = model_module.MLP(input_dim=25, hidden_dims=[128, 64, 32], dropout_rate=0.3)
model.load_state_dict(torch.load(model_path, map_location=torch.device('cpu')))
model.eval()

# Load scaler
scaler = joblib.load(scaler_path)

# Make prediction
# X_new should be a pandas DataFrame with all 25 features including player KPIs
X_scaled = scaler.transform(X_new)
X_tensor = torch.FloatTensor(X_scaled)
with torch.no_grad():
    xg_prediction = model(X_tensor).numpy()

Data Requirements

To use this model, you need:

1. Shot event data with geometric and contextual features
2. Player KPI data calculated from historical player performance (shots_per_90, goals_per_90)
3. The data should be merged on player_id before making predictions

Training

The model was trained on football shot event data merged with player KPIs:

• Binary Cross Entropy loss
• Adam optimizer (lr=0.001, weight_decay=1e-5)
• ReduceLROnPlateau scheduler
• Batch size: 256
• Epochs: 50

License

MIT