Context matters for player evaluation metrics: a case study of one-on-one dribbles

BACKGROUND<br/>Valuing individual actions in soccer is focused on shots and passes, ignoring the influence of one-on-one dribbles. Current generic metrics that value individual actions, like Expected Threat (xT) [1], depend largely on the location of the action on the pitch. Successful one-on-one dribbles create opportunities to shoot or move forward, without covering a lot of distance. Thus, current approaches could undervalue one-on-one dribbles. Adding contextual information, like player locations and velocities, could improve the valuation of one-on-one dribbles.<br/>AIM<br/>We developed an xT-inspired model that includes contextual information: Dynamic Contextualized Threat (DCT). Furthermore, we aim to show the improved effectiveness of DCT in quantifying the value of one-on-one dribbles in soccer.<br/>METHODS<br/>1507 one-on-one dribbles were manually labeled into three categories: successful, neutral, and unsuccessful. xT and DCT were trained on 2 seasons of tracking and event data of the Dutch Eredivisie. xT uses only the ball location to quantify the value of the current game state. DCT is similar but exploits information on the player locations and velocities with a convolutional neural network to get a contextualized game-state representation. The one-on-one action value was calculated as the difference between the game-state value at the beginning of the dribble and at the end of the dribble. A one-way ANOVA tested if the metrics could distinguish between the labeled one-on-one dribbles. If significant, Mann-Whitney-U tests with Bonferroni correction were executed to find which individual groups differed from each other.<br/>RESULTS<br/>166 labeled one-on-one dribbles, that were not included in the model training, were used for the statistical analysis. The one-on-one action value calculated using the xT metric was not different between the groups (F(2, 164)=2.532, p&gt;0.05), but it was different when using the DCT metric (F(2, 164)=8.4525, p&lt;0.001). The posthoc analysis on the DCT-metric indicated that successful dribbles had a higher action value (0.0084) compared to neutral dribbles (0.0018; z=3.846, p&lt;0.001, r=0.64) and unsuccessful dribbles (0.0028; z=3.763, p&lt;0.001, r=0.56).<br/>CONCLUSION<br/>The DCT metric was able to distinguish between expert-labeled successful and non-successful one-on-one dribbles. Since the xT-metric was not able to do so, this highlights the importance of adding context to metrics that value actions. Although this study uses one-on-one dribbles, the DCT metric could be used for any action, including off-ball and team actions, showing great potential for player and team analysis.<br/>REFERENCES<br/>[1]K. Singh, “Introducing Expected Threat (xT).” [Online]. Available: https://karun.in/blog/expected-threat.html<br/><br/><br/><br/><br/>