We know who wins: graph-oriented approaches of passing networks for predictive football match outcomes

Football is one of the most popular sports worldwide. Football fans, in their pursuit of maximizing enjoyment from supporting their favorite teams, are eager to know the outcome of ongoing matches. However, related studies to predict football match outcomes have focused on predicting them before the match starts, using measures from previous match data. Therefore, we propose a method to predict the outcome of an ongoing match at any time during the game as one of three classes: home win, away win, or draw. We focus on the event of passes, the most fundamental unit of movement in football, and have been shown in many studies to influence match outcomes. Unlike previous studies that used generalized aggregated variables such as total pass counts for pass information, we propose a graph-based classification model that directly utilizes the dynamically changing passing network as a graph. Our proposed graph-based model effectively leverages information about the location of each pass and the players attempting these passes from the passing network. As a result, in experiments predicting the final outcome at the 45, 60, 75, and 90-minute marks, our model outperformed baseline models that used general pass measures. Furthermore, by additionally incorporating non-pass event information, our model achieved a performance improvement of 5 to 20% across various classes compared to using only the passing network. This study holds substantial implications for integrating graph theory and graph modeling into the domain of football, highlighting the impact of passing information on outcome prediction.