What to Do and Where to Go Next? Action Prediction in Soccer Using Multimodal Co-Attention Transformer

Approximately 3,000 on-ball actions occur per match in soccer, and evaluation of individual player actions in a match is essential for strategic decision support and recruitment processes. Previous studies on such evaluation have been conducted to estimate actions' values based on the probability of the next action predicted from the context of the match situation. However, while these studies on action prediction mainly focus on predicting "what to do next,'' to the best of our knowledge, there are no studies that simultaneously address the prediction of "where to go next.'' It is obvious that the value of frequent actions such as passes and dribbles varies since the subsequent match situation can alter considerably depending on its destination. Therefore, this paper proposes a novel method for predicting the next action type and its destination from the action sequence. Our method employs a co-attention Transformer to achieve effective prediction by coordinating the action as words and the tracking data, which are closely related. In experiments on StatsBomb, an open dataset of soccer match data, the proposed model outperformed baseline methods.