In soccer, attacking play refers to the process from the beginning of possession to the finishing action of the attack, which can be interpreted as the process of scoring goals. In the course of attacking play, whether or not you shoot has a lot to do...
In soccer, attacking play refers to the process from the beginning of possession to the finishing action of the attack, which can be interpreted as the process of scoring goals. In the course of attacking play, whether or not you shoot has a lot to do with scoring, and analyzing how the players played in the process is understood in the same context as describing the type of team, tactical characteristics, etc. However, a mutually organic, continuous soccer game requires a complex and comprehensive analysis that analyzes the variables in terms of individual attacking plays, and analyzes the attacking plays in units of variables. The algorithm of machine learning is trained to find patterns and correlations in the longitudinal data sets collected sequentially, and to make optimal decisions and predictions based on the analysis, enabling complex and comprehensive analysis to some extent.
This study is a study to develop and apply a shooting prediction model based on machine learning through variables that occur in the process of attacking play in Pro Football 2022 K League 1. In order to achieve the purpose of the study, three research questions were set: analyzing the characteristics of variables that occur in the process of attack play, developing a shooting prediction model using the analyzed variables, and applying and analyzing type-specific data to the developed shooting prediction model.
For the data collection for the development and application of the shooting prediction model, variables were collected on the offensive play process excluding set plays and contested situations for a total of 60 1~10R matches of the 2022 K League 1.
Through the analysis of the characteristics of the variables that occur in the process of attacking play, which is the first research problem, the attacking play leading to shooting in the K League 1 was confirmed to have more shots in the central area than in the side area by stealing the ball from a high position and advancing with a fast attack. The teams with the most differences in the starting area (x) of the attack are Incheon, Daegu, and Seoul, in that order, while the teams with the most differences in the forward distance are Suwon FC, Seongnam, and Suwon Samsung.
In the development of a shooting prediction model using the attack play variable, which is the second research problem, the performance of the initial shooting prediction model was compared through XGBoost, a decision tree, and logistic regression analysis, which is a representative algorithm of machine learning, and the best XGBoost algorithm was selected, and the final shooting prediction model was developed through hyperparameter adjustment, and as a result of performance evaluation, the accuracy was 98%, the accuracy was 98%, the reproduction was 98%, and the F1-score was 98%. Through the importance of the variables, the main variables applied to the prediction were identified in the following order: attack end angle, forward distance, attack start position (x), attack speed, and attack start position (y).
The third research problem, the developed shooting prediction model, applied type-specific data to the developed shooting prediction model, and the analysis applied data to the developed final model according to the type of play style and team performance type, and the applied data was compared with the FN and FP values that failed to predict and the values that succeeded in predicting shooting.
The variables that occur in the course of attacking play presented in this study explain the content of the game and provide basic data for game analysis and performance improvement. In addition, it suggests that the shooting prediction model can not only improve performance by identifying the type and tactical characteristics of the team, but also use it as a tool for situational feedback through FN and FP values based on the usability of the shooting prediction model.