Seq2Event: Learning the Language of Soccer Using Transformer-based Match Event Prediction

Abstract

Soccer is a sport characterised by open and dynamic play, with player actions and roles aligned according to team strategies simultaneously and at multiple temporal scales with high spatial freedom. This complexity presents an analytics challenge, which to date has largely been solved by decomposing the game according to specific criteria to analyse specific problems. We propose a more holistic approach, utilising Transformer or RNN components in the novel Seq2Event model, in which the next match event is predicted given prior match events and context. We show metric creation using a general purpose context-aware model as a deployable practical application, and demonstrate development of the poss-util metric using a Seq2Event model. Summarising the expectation of key attacking events (shot, cross) during each possession, our metric is shown to correlate over matches (𝑟 = 0.91, 𝑛 = 190) with the popular xG metric. Example practical application of poss-util to analyse behaviour over possessions and matches is made. Potential in sports with stronger sequentiality, such as rugby union, is discussed. CCS CONCEPTS• Applied computing; • Computing methodologies → Machine learning; Model development and analysis;

Transformer-Based Neural Marked Spatio Temporal Point Process Model for Football Match Events Analysis

“…The feasible solution from the NTPP framework [23] is to encode the information from the history events information ( y 1 , y 2 , ... y i−1 ), y i = [t i , m i , z i ] into a fixed-size single vector h i with LSTM [10], GRU [3], or Transformer Encoder [31]. Based on a previous study [25], Transformer Encoder has been found to be slightly less effective, but significantly more efficient than LSTM. Therefore, in this study, we applied Transformer Encoder to encode the history events information.…”

“…Stage 2: History encoding. In this stage, a dense layer is first applied to interevent time t i and other continuous features, with an embedding layer applied to zone z i and action m i respectively, allowing the model to better capture information in the features [25]. Afterward, with the position encoding and transformer encoder from the Transformer model [31] (more details on Appendix D), a fixed-size encoded history vector with size (31) can be retrieved.…”

“…Data analysis can be leveraged by players to improve performance, and generate insight for the coaching process and tactical decision-making. Furthermore, it provides fans with quantified measures and deeper insights into the game [2,25]. For years, analysis and research have focused on players' actions when they are in possession of the ball, with each player statistically shown to have 3 minutes on average with the ball [9].…”

“…There have been numerous attempts in the existing sports analysis literature to understand how the sequence of events in the past affects the next event or its outcome. The majority of the studies [21,24,25,32,35] used machine learning (ML) models to handle the long sequences of event data. The ML models encoded the long history events vectors (each vector representing a historical event, and with the event being described by features in the vector) into a vector representation by incorporating Long Short-Term Memory (LSTM) [10], Gated Recurrent Unit (GRU) [3], or Transformer Encoder [31].…”

“…The Seq2Event model was recently proposed by Simpson et al [25]. The model combines the encoding method and dense layers to forecast the location and type of the next event.…”