Detecting Match State Changes in Ice Hockey From Positional Data
Ryttyläinen, Juha-Matti (2021)
Ryttyläinen, Juha-Matti
2021
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2021-12-07
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202111188489
https://urn.fi/URN:NBN:fi:tuni-202111188489
Tiivistelmä
This thesis aims to determine whether it is possible to utilize machine learning techniques to determine whether an ice hockey match is in a state of active play or not. As an additional requirement, the match state determination should work in real-time. To determine the match state, the method demonstrated in this thesis utilizes position data of players and pucks produced by the Wisehockey platform. The dataset used in the presented experiment consisted of 40 professional ice hockey matches containing the position data for players and pucks. In addition to the position data, the dataset contains all the match state information for the matches, serving as ground truth for machine learning.
The technical problem presented was approached as a classification problem of small time windows of position data. The machine learning model chosen for the experiment was the random forest classifier. The model achieved AUC scores of above 0.89 for all matches in a 10 match test set. However, the method was not robust enough for a real world application but manages to provide insight into the technical problem discussed. From the results in this thesis it seems possible that by improving the method, it is possible to determine the match state accurately enough for real world use. On the other hand, using a different machine learning method entirely might also produce good results.
The technical problem presented was approached as a classification problem of small time windows of position data. The machine learning model chosen for the experiment was the random forest classifier. The model achieved AUC scores of above 0.89 for all matches in a 10 match test set. However, the method was not robust enough for a real world application but manages to provide insight into the technical problem discussed. From the results in this thesis it seems possible that by improving the method, it is possible to determine the match state accurately enough for real world use. On the other hand, using a different machine learning method entirely might also produce good results.