Authors:
Hikaru Yoshihara
1
;
Ning Ding
2
and
Keisuke Fujii
1
Affiliations:
1
Nagoya University, Japan
;
2
Nagoya Institute of Technology, Japan
Keyword(s):
Machine Learning, Volleyball, Tracking Data.
Abstract:
In volleyball, statistical analysis based on data aggregation at the team or match levels has developed, and its use for player performance evaluation and tactical analysis has expanded. However, there has been limited discussion on the quantitative evaluation of how individual plays affect rally outcomes. To address this issue, a model that predicts rally outcomes under specific conditions using player location data is useful. This study aims to evaluate plays based on a prediction model, focusing on the first transition following a back-row attack. We extracted 103 target scenes from game footage recorded from behind the end line and manually created tracking data for six players per team. Using this dataset, we trained an XGBoost model to predict the future probability of scoring and the probability of blocking by two or more opponents in each game state (receive, toss, attack). To quantify play evaluation, we propose the Valuating Volleyball States by Estimating Probabilities (V2
SEP), which expresses play evaluation values in each state based on the prediction model, weighting them according to the percentage of points scored when a player is blocked. To verify the validity of the prediction model used in V2SEP, we assessed F1 scores and SHAP values for each state. The results indicate that the predictions were reasonably accurate and reflected not only the contributions of directly involved players but also those of other players affecting scoring and block induction. Furthermore, the play evaluation metrics demonstrate expected trends whereas some scenes show the limitations, suggesting that V2SEP may be useful for play evaluation in volleyball.
(More)