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

Abstract

With recently available football match event data that record the details of football matches, analysts and researchers have a great opportunity to develop new performance metrics, gain insight, and evaluate key performance. However, most sports sequential events modeling methods and performance metrics approaches could be incomprehensive in dealing with such large-scale spatiotemporal data (in particular, temporal process), thereby necessitating a more comprehensive spatiotemporal model and a holistic performance metric. To this end, we proposed the Transformer-Based Neural Marked Spatio Temporal Point Process (NM-STPP) model for football event data based on the neural temporal point processes (NTPP) framework. In the experiments, our model outperformed the prediction performance of the baseline models. Furthermore, we proposed the holistic possession utilization score (HPUS) metric for a more comprehensive football possession analysis. For verification, we examined the relationship with football teams' final ranking, average goal score, and average xG over a season. It was observed that the average HPUS showed significant correlations regardless of not using goal and details of shot information. Furthermore, we show HPUS examples in analyzing possessions, matches, and between matches.

Action Valuation of On- and Off-Ball Soccer Players Based on Multi-Agent Deep Reinforcement Learning

“…Numerous approaches have been employed to quantitatively assess the actions of attacking players in terms of scoring, such as using expected scores derived from tracking data [7], [15]- [18], and action data such as dribbling and passing [1], [4], [19]. Some researchers have valuated passes [20]- [22], while others have assessed actions to receive a ball by attributing a value to the location with the highest expected score [10], [23] and using a rule-based approach [24].…”

“…In particular, each action's value has been quantified based on the strength of its association with scores for on-ball players (i.e., with a ball). In previous work using supervised learning, machine learning models were used to compute an action's value by predicting whether scores or other events occur or not in the following actions [1]- [4]. In these frameworks, it would be difficult to consider possible (i.e., counterfactual) actions as time goes back from a goal or other events.…”

Masked modeling-based action event prediction considering bidirectional time-series in soccer

“…1 Specifically in soccer, researchers have focused on assessing the effectiveness of each player's decision-making and team strategy, with recent attention also drawn to machine learning-based methods. [2][3][4][5][6] Conventional methods evaluate the quality of a player's decision-making by predicting the occurrence probability of the next or subsequent important action event based on the current match situation. These prediction-based methods successfully evaluate the individual action levels although it is difficult to interpret from the stats such as ball possession or number of shots compiled throughout the entire match.…”

“…: angle from the center of the opposition goal 6] : current score advantage Note that the above continuous features are normalized before inputting the fully connected (FC) layer and concatenated to token embeddings. The concatenated embeddings are input to the Transformer encoder along with the positional embeddings E pos t as follows:…”