3 67 MACHINE LEARNING SUPPORTING SUBSTITUTIONS IN SOCCER 4. DISCUSSION The main goal of this study was to explore the possibility of predicting on physical performance of individual players and decision support for coaches to help them make an informed decision on player substitutions. Our study focused on a player’s physical performance within the match, making the identification of underperforming players the critical point. In line with previous research, this study revealed that entire match players show a significant decline in physical performance during the match in distance covered, distance covered in speed category, and energy expenditure in power category [4], [7]. While earlier studies found a decline of a 10-15% reduction of the HIR and VHIR from the first to the second half [2], [3], our results did not show any decline in these high-intensity type-1 variables. Thereby, our findings are in consent with more recent studies [26], [27]. Furthermore, our results replicate the study of Liu et al.[26], who found that time spent in the very low intensity (VLIR) category is increasing while time in medium intensity categories is decreasing (LIR and MIR) and time in high-intensity categories are stable throughout a match. The same pattern can be seen for the energy expenditure in different power categories. Given these results, we can support our first hypothesis that type-1 and type-2 load variables can identify decreasing player performance throughout a match. In order to answer our second research question, we found that substitutes perform better than entire match players on both type-1 and type-2 variables. Most of the substitutions occur at halftime and during the 60-90 minute period, which aligns with previous research [27]. In agreement with the literature, substitutes who had been introduced during the second half covered more distance and performed more high-intensity activities relative to entire match players over the same period [8]. In addition, second-half substitutes spent more energy in higher power categories [28]. As substitutes demonstrate higher values in physical performance variables than the entire match players, the substitution of underperformers may improve the team’s performance and make the difference between winning and losing [5]. This study’s machine learning models can identify and predict a players’ physical performance in an early stage of the match. The Random Forest model outperformed both the Decision Tree and Naïve Bayes algorithm. For every threshold, the Random Forest model identified the underperformers and performers best. The precision of the variable energy expenditure in power category outperformed models based on the variables distance covered and distance covered in speed category. The outperformance of the variable energy expenditure in power category illustrates that the more advanced type2 variable is most sensitive to recognizing a player’s physical performance in an early stage of the match. The stronger the relation in reality between the variable and the outcome, the higher the precision of the machine learning model may be expected [29].
RkJQdWJsaXNoZXIy MjY0ODMw