This repository contains the Pytorch implementation code for the paper "We Know Who Wins: Graph-oriented Approaches of Passing Networks for Predictive Football Match Outcomes"
Both the passing graph and in-game features of each team are input to the model. The passing graph passes through three graph network blocks consisting of graph attention-elu-dropout layer to be pooled into graph embedding. The graph embeddings of both teams that have been fused with the in-game features are combined again to finally return the probability for each class.
Please refer to requirements.txt for package dependency to reproduce the code in this repository.
The datasets are available from the corresponding author upon reasonable request.
You can run every steps from preprocessing the raw html data to testing new example through the following commands.
- Preprocessing
This step takes raw html files as input and returns preprocessed torch_geometric.Data-type json file. The pred_min argument refers to the time period you want to make passing network to predict the outcome of the match. The default value will be set as 90 for pred_min argument unless provided.
python3 preprocess.py --gpu_num [YOUR GPU NUM] --pred_min [Prediction Minute]However, we highly recommend you to skip the preprocessing part and directly go to the training part utlizing the uploaded preprocessed json files final_home_90.json and final_away_90.json, and the scaler scaler.json since this step might take more than an hour depending on the gpu capability
- Training
This step takes two json files (Home Data and Away Data) and trains the proposed model in our study. You can run this step through the following command.
python3 training.py --gpu_num [YOUR GPU NUM]The hyperparameters needed to train the model will be set as default as 1000 epochs with 10 patiences, 0.001 of learning rate and 90 minutesof prediction minute, unless you designate them by arguments. Plus, if you want to train the model for early-stage prediction like 45-minute or 70-minute, designate the value by providing additional argument --pred_min at the end of the command as below.
python3 training.py --gpu_num [YOUR GPU NUM] --n_epochs 500 --e_patience 100 --lr 0.01 --pred_min 45- Testing
With the trained model, you can test the model performance. Here, there are two options you can take depending on whether you want to test the model using the testloaders made after preprocessing and trainig steps of this repository or to test the model with the whole new match data.
For the former, you can do it through the following command
python3 testing.py --use_testloader yes --new_scaler no --gpu_num [YOUR GPU NUM]For the latter, you can do it through the following command
python3 testing.py --use_testloader no --new_scaler no --gpu_num [YOUR GPU NUM]Here, the --new_scaler argument is [no] when you want to test a single match, or [yes] when you want to test more than one matches.