Off-puck scoring opportunities: Detection and visualization of ice-hockey scoring opportunities from position data
Mäenpää, Tuomas (2024)
2024
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
2024-03-25
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202402222454
https://urn.fi/URN:NBN:fi:tuni-202402222454
Tiivistelmä
The purpose of this thesis was to investigate if off-puck player positions can be utilized to predict goal-scoring in ice hockey. In addition to predicting scoring, the aim was to predict where the goals would be scored from. Scoring goals is the most important aspect of any ball-sport. In ice hockey, players spend most of the match without the puck, yet they continuously impact the course of the match. During attacking play, the players without the puck support the controlling player with their movement and positioning. Currently, there are no methods to quantify the positioning of offpuck players during attacking play. The objective of this thesis was to develop a method to detect and locate off-puck scoring opportunities based on position data.
The data used in this study was collected with Wisehockey’s sports tracking system. The dataset contained all shot events and all successful pass events from 250 professional ice hockey matches during the 2022/2023 Liiga season. In addition to the events, full position data for the puck and players was available from those matches.
The Off-Puck Scoring Opportunity (OPSO) model was built from three separate probability models. The models estimated the probability of pitch control, the probability of a successful pass, and the probability of scoring. The output of the OPSO model was a probability density map that represented the probability of scoring within 5 seconds from the next on-puck event from any location within the rink. The pitch control probability was modeled with a parametric approach based on the position data of the players. The pass probability was modeled by assimilating the displacement of the puck between consecutive events. Three machine-learning methods were used to estimate the probability of scoring: exponential distribution, logistic regression, and a multilayer perceptron. The models were tasked to predict goal-scoring based on event location. The classifiers reached AUC scores between 0.81 and 0.83. The objective was to exceed the prediction quality with OPSO.
Two validation methods were used to test OPSO. First, a convolutional neural network (CNN) was trained and tested on the output heatmaps produced by OPSO. The CNN reached the AUC score of 0.72. Second, the total scoring probabilities were integrated from the output probability density maps. The AUC score for the integrated probabilities from OPSO was 0.84. The results showed that OPSO improved the ability to detect off-puck scoring opportunities. Taking players’ positioning and movement into account in scoring probability estimation enhanced the prediction reliability. The objective of the thesis was met and the purpose was fulfilled.
The data used in this study was collected with Wisehockey’s sports tracking system. The dataset contained all shot events and all successful pass events from 250 professional ice hockey matches during the 2022/2023 Liiga season. In addition to the events, full position data for the puck and players was available from those matches.
The Off-Puck Scoring Opportunity (OPSO) model was built from three separate probability models. The models estimated the probability of pitch control, the probability of a successful pass, and the probability of scoring. The output of the OPSO model was a probability density map that represented the probability of scoring within 5 seconds from the next on-puck event from any location within the rink. The pitch control probability was modeled with a parametric approach based on the position data of the players. The pass probability was modeled by assimilating the displacement of the puck between consecutive events. Three machine-learning methods were used to estimate the probability of scoring: exponential distribution, logistic regression, and a multilayer perceptron. The models were tasked to predict goal-scoring based on event location. The classifiers reached AUC scores between 0.81 and 0.83. The objective was to exceed the prediction quality with OPSO.
Two validation methods were used to test OPSO. First, a convolutional neural network (CNN) was trained and tested on the output heatmaps produced by OPSO. The CNN reached the AUC score of 0.72. Second, the total scoring probabilities were integrated from the output probability density maps. The AUC score for the integrated probabilities from OPSO was 0.84. The results showed that OPSO improved the ability to detect off-puck scoring opportunities. Taking players’ positioning and movement into account in scoring probability estimation enhanced the prediction reliability. The objective of the thesis was met and the purpose was fulfilled.