Measuring defensive quality in soccer
The art of good defending is to prevent something from happening before it has even happened. Virgil Van Dijk is considered one of the best defenders in world soccer as he has the ability to prevent a pass being made to an open attacker to shoot by forcing the ball carrier to pass somewhere else less dangerous. However,while we know this is great defending, in today’s stats, Van Dijk would not receive any acknowledgement. A defender’s contribution is simply measured by the number of tackles or interceptions they make. But what if we were able to measure actions that have been prevented before they were made?
The aim of a defense and a defender is to make offensive play predictable. For example, Jürgen Klopp’s Liverpool, press the opposition with the aim of forcing them to give the ball away in specific areas of the pitch by limiting the number of passing options available in dangerous areas. If the art of good defending is to make play predictable, then it should be measurable. Given enough data, we should be able to predict where a player will pass the ball, the likelihood of that pass being completed and whether this pass will result in a scoring opportunity. It therefore stands that we should be able to measure if a defender forces an attacker to change their mind or to prevent an attacker from even becoming an option.
Figure 1 shows a situation from a match between Liverpool vs Bayern Munich in the 2018/19 UEFA Champions League that leads to Mané (red 10) scoring. Our model identifies that Milner (red 7) is the primary target for Van Dijk (red 4) in the first instance. However, due to the combination of Gnabry (blue 22) closing down Milner, Lewandowski (blue 9) closing down Van Dijk and Mané making an active run, behind the defence, Mané becomes both the most likely receiver and a high threat for scoring. This demonstrates our ability to model how players decision making is influenced and how a situation can move from low threat to high threat by the off-ball actions of attackers and defenders.
In this paper we present a novel Graph Convolutional Neural Network (GNN) which is able to deal with highly unstructured and variable tracking data to make predictions in real time. This allows us to accurately model
defensive behaviour and its effect on attacking behaviour, i.e., preventing actions before they have occurred.