Repository files navigation

Source code for the paperSoccerCPD: Formation and Role Change-Point Detection in Soccer Matches Using Spatiotemporal Tracking Data by Kim et al., KDD 2022.

SoccerCPD is a change-point detection framework that distinguishes tactically intended formation and role changes from temporary changes in football matches using spatiotemporal tracking data.
It first assigns roles to players frame-by-frame based on the role representation (Bialkowski et al., 2014) and performs two-step nonparametric change-point detections (Song and Chen, 2020): (1) formation change-point detection (FormCPD) based on the sequence of role-adjacency matrices and (2) role change-point detection (RoleCPD) based on the sequence of role permutations.

Here is an example of applying SoccerCPD to a match. It shows that the match is split into four formation periods (where the team formation is consistent) and five role periods (where the player-role assignment remains constant).

We cannot share the entire dataset due to the security issue, but every process except for the formation clustering can be reproduced using the sample match data (17985.ugp) since applying our method to a match does not require data from other matches. For the formation clustering (Section 4.3) and role labeling (Section 5.3) steps of the paper, we offerdata/form_periods.pkl,data/role_periods.csv, anddata/role_records.csv that contains the information of all the detected formations and roles in our dataset.

We have implemented the role representation algorithm using Python on our own, while adopted the R packagegSeg for discrete g-segmentation. Therefore,both Python and R need to be installed for executing the code. The version we have used in this study are as follows:

• Python 3.8
• R 3.6.0

To perform the whole process at once, we utilize the Python packagerpy2 to run the R script withgSeg inside our Python implementation. We found thatthe M1 chip by Apple does not support binaries forrpy2's API mode in Python, raising a memory error. (So please use another processor such as the Intel chip.)

After installing the necessary languages, you need to install the packages listed inrequirements.txt. Make sure that you are in the correct working directory that contains ourrequirements.txt.

pip install -r requirements.txt

Subsequently, please download the sample match data (named17985.ugp) from the following Google Drive link and move it into the directorydata/ugp.

• Link for the data:https://bit.ly/3gHhkHG

Finally, you can run the algorithm for the sample match data simply by executingmain.py.

python -m main

For further analyses, opentutorial.ipynb and run the cells in order from the above. (Section 4.1, 4.2, 5.1, and 5.2)

You can also reproduce the results described in the paper by executing the following notebooks:

• notebooks/1_formaiont_clustering.ipynb (Section 4.3)
• notebooks/2_role_labeling.ipynb (Section 5.3)
• notebooks/3_model_evaluation.ipynb (Section 6.1)
• notebooks/5_switching_pattern_discovery.ipynb (Section 6.2)
  

Lastly, we visualize our results as animations using Tableau 2020.4. The full-version video is available inhttps://www.youtube.com/watch?v=F-tvG1-MLps with the description in Appendix C of the paper.

This repository includes the source code for the following paper and tracking data collected from a sample match (17985.ugp). Please cite the paper when using our code or the sample match data.

@inproceedings{Kim2022,  author       = {Kim, Hyunsung and                  Kim, Bit and                  Chung, Dongwook and                  Yoon, Jinsung and                  Ko, Sang{-}Ki},  title        = {{SoccerCPD}: Formation and Role Change-Point Detection in Soccer Matches  Using Spatiotemporal Tracking Data},  booktitle    = {The 28th {ACM} {SIGKDD} Conference on Knowledge Discovery and Data Mining},  pages       = {3146--3156},  publisher    = {{ACM}},  year         = {2023},  location     = {Washington, DC, USA}  isbn         = {979-1-4503-9385-0}  url          = {https://doi.org/10.1145/3534678.3539150},  doi          = {10.1145/3534678.3539150},}

• A. Bialkowski, P. Lucey, P. Carr, Y. Yue, S. Sridharan and I. Matthews, Large-Scale Analysis of Soccer Matches Using Spatiotemporal Tracking Data, IEEE International Conference on Data Mining, 2014, DOI:https://doi.org/10.1109/ICDM.2014.133.
• H. Song and H. Chen, Asymptotic distribution-free changepoint detection for data with repeated observations, Biometrika, 2021, asab048, DOI:https://doi.org/10.1093/biomet/asab048.