Enhancing Formation Analysis Using

Multi-layer Graph Convolutional Neural Networks

and Geometric Formation Features
2024 Joint 13th International Conference on Soft Computing and Intelligent Systems and 25th International Symposium on Advanced Intelligent Systems (SCIS&ISIS) | 979-8-3503-7333-2/24/$31.00 ©2024 IEEE | DOI: 10.1109/SCISISIS61014.2024.10760047

Shunsuke Takagi Yohei Nakada

Graduate School of Advanced Mathematical Sciences Graduate School of Advanced Mathematical Sciences
Meiji University Meiji University
Tokyo, Japan Tokyo, Japan
yoheinakada@[Link]

Abstract—We have previously proposed a formation analysis method incorporating the proposed layers by applying it to a
method using multi-layer graph convolutional neural networks soccer shooting scene classification problem based on that in
with geometric formation features as input. In addition, this [19]. We also conduct an ablation study on the adjacency
analysis method has been validated by applying it to a information input to investigate its impact on performance.
classification problem of soccer shooting scenes, and the results
have been demonstrated its validity. However, when considering
the context of team sports, it was considered that there is room for II. RELATED WORK
enhancement in the graph convolutional layers used in the method. To date, numerous studies have been conducted on team
Therefore, in this study, we propose new graph convolutional sports, focusing on methods, software, devices, and systems that
layers that are considered more effective for team sports. address various aspects of performance, strategy, and player
Furthermore, we validate the analysis method incorporating the behavior, as many examples reviewed and/or explained in [22]–
proposed layers by applying it to a soccer shooting scene [34]. Therefore, this section introduces studies that are closely
classification problem. We also conduct an ablation study on the related to our study.
adjacency information input.
In references [1]–[4], considerable contributions have been
Keywords—sports data analysis, geometric formation features, made in the field of geometric formation analysis. These studies
multi-layer graph convolutional neural networks, graph primarily utilize Voronoi areas or their extensions to depict
classification problem, soccer shoot scene classification players’ dominant areas. For example, reference [1] extends
Voronoi areas with a player’s motion model to evaluate
I. INTRODUCTION teamwork in soccer. Reference [2] adopts a more natural motion
model to further improve the extended areas, applying these to
In recent years, there has been a notable increase in the
analyze teamwork in field hockey. Meanwhile, standard
collection and utilization of player and ball tracking data in
Voronoi areas are also used in [3] to capture players’ spatial
various team sports. This trend is particularly pronounced in
dynamics in Futsal games. Reference [4] introduces soccer
invasion team sports such as soccer, basketball, rugby, and
analyses via standard Voronoi areas and their adjacency
American football. Concurrently, for invasion team sports, there
information, namely Delaunay graphs.
has been substantial development of approaches related to
geometric formation analysis [1]–[9] using players’ dominant Several studies [5]–[9] have introduced various methods
areas or the adjacency information related to these areas. employing geometric formation features derived from these
Furthermore, there has been a growing interest in applying geometric formation analyses. Concretely, reference [5] presents
multi-layer neural networks to the analysis of tracking data in methods to identify key players and similar scenes via features
invasion team sports [10]–[18]. However, despite the potential from adjacency information related to the extended Voronoi
benefits of combining geometric formation analysis with multi- areas in [2], applying them to a real soccer match. Reference [6]
layer neural networks, to the best of our knowledge, no study has discusses a method for classifying soccer passes using features
yet taken on such a challenge. of the extended Voronoi area in [1]. Reference [7] introduces a
clustering method for soccer formations using features of
On such background, in our previous work [19], we proposed
Delaunay graphs. Additionally, references [8] and [9] describe
a formation analysis method using multi-layer graph
methods for quantifying soccer pass plays employing features of
convolutional neural networks with geometric formation
both the extended Voronoi areas in [2] and their adjacency
features as input. In addition, this analysis method has been
information.
validated by applying it to a classification problem of soccer
shooting scenes, and the results have been demonstrated its Moreover, recent advancements have led to the development
validity. However, when considering the context of team sports, of various methods utilizing multi-layer neural networks such as
it was considered that there is room for enhancement in the graph convolutional neural networks, generative adversarial networks,
convolution layers [20][21] used in our previous works. long short-term memory networks, graph neural networks, and
Therefore, in this study, we propose new graph convolutional their combinations for analyzing tracking data in team sports
layers which allow both positive and negative edge weights for [10]–[18]. Specifically, reference [10] explains a generative
formation analysis. Furthermore, we validate the analysis model to capture human behaviors, demonstrating its

Authorized licensed use limited to: INSTITUTO FEDERAL DE SAO PAULO. Downloaded on May 07,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
 TABLE I. NODE FEATURES ADOPTED IN [19] formation analysis. Table I presents the specific node features
# Features adopted in the experiments in our previous study [19]. These
1 Team category (Right team: -1, Left team: 1)
were also used in the experiment described later in this paper.
2 Player category (Field player: 0, Goalkeeper: 1) B. Structure of Multi-layer Convolutional Neural Networks
3 x-coordinate location (horizontal) As mentioned earlier, in the target method [19], multi-layer
4 y-coordinate location (vertical) graph convolutional neural networks, which are relatively new
5 x-coordinate directional speed types of graph neural networks [35][36], have been adopted.
6 y-coordinate directional speed These types of networks consist of multiple graph convolution
7 Relative x-coordinate location to the ball layers that execute convolutions based on graph structures
8 Relative y-coordinate location to the ball determined by the aforementioned adjacency information. For
9 Relative x-coordinate directional speed to the ball the application to the soccer shooting scene classification
10 Relative y-coordinate directional speed to the ball problem in [19], after the last convolution layer, the features of
11 Dominant area ratio (with the whole field area set as 1) all nodes are aggregated within an aggregation layer. Then, the
Dominant area ratio inside the opponent's side area (with the half-field shooting scene category is predicted using aggregated features.
12 A significant advantage of employing such graph neural
area set as 1)
13
Dominant area ratio inside the opponent's penalty area (with the whole networks [22][23] is their capability to remain permutation-
penalty area set as 1) invariant and/or equivariant under certain conditions, obviating
concerns regarding the order of players within data. Fig. 1
effectiveness with basketball tracking data. Reference [11] illustrates a schematic of such a structure in the target method
presents another generative model for team sports to predict [19] for the shooting scene classification problem.
players’ behavior in possible future scenarios. Reference [12]
introduces a generative model to simulate defensive players’
movements given offensive players’ movements. References IV. ENHANCING FORMATION ANALYSIS METHOD
[14] and [15] describe other types of models to predict basketball This section describes an enhancement of the
player movements. References [15] and [17] discuss predictive aforementioned target method in this study. Concretely, we first
models to identify potential soccer pass receivers using video describe the two kinds of graph convolution layers adopted in
data and tracking data. Reference [16] describes a model to our previous study [19]. Then, we discuss the room for
identify the roles of defensive players from tracking data in enhancement in these graph convolution layers, considering the
soccer corner kick scenes. Reference [18] also introduces a context of team sports. Finally, we introduce the new
predictive model focusing on soccer corner kick scenes.
convolution layers proposed in this study for enhancement.
As mentioned in this section, both geometric formation
A. Graph Convolution Layers Adopted in [19]
analysis and multi-layer neural networks have been applied in
team sports. Therefore, there seem to be high potential benefits In the experiment in our previous study [19], two different
of combining geometric formation analysis and multi-layer kinds of graph convolution layers [20][21] were adopted and
neural networks. However, to the best of our knowledge, we their performances were compared. The first one was the
could not find any study that has taken on such a challenge standard graph convolutional layer [20], described by the
except for our previous study [19] and this study. following equation:
() () () () ()
ℎ = ℎ + , ℎ . (1)
III. TARGET FORMATION ANALYSIS METHOD ∈ ()
() ()
In this section, we outline the formation analysis method Here, the matrices and represent the weight
proposed in our previous study [19], which is targeted for parameters used for the convolutional operation in the -th layer.
()
enhancement in this study. Specifically, we first present the The vector ℎ denotes the input feature vector for the -th node
geometric formation features that serve as inputs in this method ()
[19]. We then describe the structure of the multi-layer graph at the -th layer, and the vector ℎ indicates the output feature
convolutional neural networks used in this method [19]. vector (before applying an activation function) for the -th node
at the same layer. The function ( ) returns the set of the
A. Geometric Formation Features neighboring nodes of the -th node. The variable , symbolizes
In the target method [19], the geometric formation features the weight of the edge from the -th node to the -th node,
serve as inputs, derived from players’ dominant areas (i.e., typically set to 1.
Voronoi areas or their extensions) and their adjacency The second one was the graph transformer convolution
information. More concretely, we adopted the extended Voronoi layer [21]. With this layer, the convolutional operation is
areas as described in [2] as the players’ dominant areas as they described in the same way as in (1); however, the edge weight
appeared to be the most natural choice based on our investigation.
, is calculated by employing a part of the transformer’s
Their adjacency information was then employed to define the
structure [37] as follows:
graph structure of multi-layer graph neural networks, where the
nodes represent individual players. Additionally, the nodes were exp ,
characterized by features, including those derived from the , = , (2)
players’ dominant areas as well as variables (e.g., locations, ∑ ∈ ( ) exp ,
velocities, and team categories), which are essential in soccer

Authorized licensed use limited to: INSTITUTO FEDERAL DE SAO PAULO. Downloaded on May 07,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
Fig. 1. Schematic of formation analysis method [19] for shooting scene classification problem.
6
! sinh
, = , (3) , =
,
. (5)
√$ ∑ ∈ ( )0sinh , 0
() () () ()
= 3 ℎ , ! = 4 ℎ . (4) Obviously, this layer is directly inspired by the edge weight (2)
of the graph transformer convolution layer [21], by expanding to
Here, the vector expresses the query vector of the -th node, include negative edge weights.
and the vector ! is the key vector of the -th node. The matrices
() () The second and third types replace the hyperbolic sine
" and # indicate the weight parameters used for the function used in the computation (5). Specifically, the second
calculation of these vectors in the -th layer. The natural number type adopts a simple linear function as follows, which exhibits a
$ signifies their dimension, which is also identical to the more gradual increase in the edge weight compared to the
()
dimension of the vector ℎ . hyperbolic sine function:
,
B. Room for Enhancement in Target Method [19] , = . (6)
∑ ∈ ( )0 , 0
In the experiment in our previous study [19], the
aforementioned two different graph convolution layers [20][21] In this layer, extreme differences in the absolute values of the
edge weights , are unlikely to occur because there is no
were adopted in the method [19], and their performances were
exponential function. Therefore, it is expected to have the
compared. Subsequently, the results demonstrated their
characteristic of perceiving the importance of both teammates
reasonable validity; however, there was room for further
and opponents around target players more evenly, which may be
enhancement when considering the context of team sports. effective for formation analysis in team sports.
Specifically, despite it being intuitive to consider that edges
from opponents and those from teammates have edge weights The third type further enhances this characteristic by
, with opposite signs, edge weights in such layers can take
employing the hyperbolic tangent function tanh()), which has
only positive values. Therefore, in this study, we introduce saturation characteristics, as the following equation:
tanh ,
(7)
another convolution layers based on the graph transformer = .
,
convolution layer [21], which allow the edge weights , to take ∑ ∈ ( )0tanh , 0
both positive and negative values, as described in the next It should be noted that these graph convolution layers (5)–
subsection. (7) cannot be achieved by merely shifting the range of edge
weights , in (2).
C. Enhanced Graph Convolution Layers
Specifically, we introduce three types of graph convolution V. VALIDATION EXPERIMENT
layers to overcome the issue mentioned in the previous In this experiment, as in our previous study [19], we applied
subsection. The first type is an enhanced graph transformer the formation analysis method [19] and its enhanced methods in
convolution layer utilizing the hyperbolic sine function: this study to a shooting scene classification problem. In this
+,-(.)/ +,-(/.)
sinh()) ≔ , where the edge weight , is computed problem, the tracking data for each frame is classified into three
by the following equation: categories: left-team shooting scene, right-team shooting scene,
or non-shooting scene. Based on such a classification problem,

Authorized licensed use limited to: INSTITUTO FEDERAL DE SAO PAULO. Downloaded on May 07,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
it is possible to compute a dominance index of the teams, which TABLE II. VALIDATION ACCURACIES (%)
WITH FIVE TYPES OF GRAPH CONVOLUTION LAYERS
can serve as additional explanatory information in broadcasting.
Acc. of Average Acc. of
A. Data Preparation Type of Graph Convolution Layers
Best Config 5-best Configs.
We utilized tracking data from 45 matches sourced from Standard Graph Convolution [20] 92.452 92.290
DataStadium Inc. [38], as in our previous study [19]. In the Graph Transformer Convolution [21] 92.253 92.161
previous study, we divided 45 matches into 30 matches for Ours (sinh) 92.623 92.524
training, 10 for validation, and 5 for testing. However, for more Ours (linear) 92.675 92.544
stability, we divided it into 25 for training, 10 for validation, and
Ours (tanh) 92.611 92.515
10 for testing in this study. Also as in the study [19], in this
experiment, we extracted one-second intervals (25 frames) TABLE III. TEST ACCURACIES (%)
before each shooting in each match, referring to relevant play- WITH FIVE TYPES OF GRAPH CONVOLUTION LAYERS
log data as the shooting scenes. Here, only frames of exactly 22 Acc. of Average Acc. of
players were included. Additionally, considering rotational Type of Graph Convolution Layers
Best Config 5-best Configs.
symmetry, we included the frames rotated by 180 degrees as the
Standard Graph Convolution [20] 88.872 89.286
shooting scenes from the opposite direction. Moreover, we
Graph Transformer Convolution [21] 90.145 89.823
balanced each dataset by randomly adding an equivalent
number of non-shooting scenes as the number of shooting Ours (sinh) 90.112 90.048
scenes from one side. Ours (linear) 89.454 89.582
Ours (tanh) 89.986 89.672
As a result, our training, validation, and test datasets each
contained an equal number of frames for right-team shooting, Geometric library [39]. Other parameters were set to the default
left-team shooting, and non-shooting scenes. The total numbers values within the package.
in the three datasets are 48,675, 17,025, and 18,225 frames,
respectively. We then calculated the features described in Table C. Exprimental Results
I and the adjacency information for each frame. Here, we To organize the experimental results, for each type of graph
adopted the player motion model with its parameters in [9] to convolution layers, we determined the best configuration for the
compute the geometric formation features and the adjacency type as the one that recorded the highest accuracy in the
information. Each feature, except for team and player categories, validation dataset using the best-trained parameter set across all
was standardized with its mean and variance in the training configurations of the number of graph convolutional layers and
dataset. the number of hidden features. Additionally, in the same manner,
B. Settings for Graph Convolutional Neural Networks we selected the 5-best configurations for each type based on
accuracies against the validation dataset. Table II displays the
In this experiment, we compared multi-layer graph validation accuracy of the best configuration and the average
convolutional neural networks using five different types of validation accuracy of the 5-best configurations for each type of
graph convolutional layers outlined in the previous section. To graph convolution layers. Table III presents the test accuracy of
do this, we first implemented three new layers proposed in this the best configuration and the average test accuracy of the 5-
study. Accounting for the findings in our previous study [19], best configurations for each type. In these tables, the highest
all the neural networks are set with the number of graph values in all types are highlighted in bold, while the second-
convolutional layers between 2 and 5, and the number of hidden highest values in all types are indicated in italics.
features between 3 and 20. Moreover, for each configuration of
the neural networks, we executed ten trials. As indicated in these tables, the validation accuracies of the
best configurations and the average validation accuracies of the
The other settings were the same as in our previous study 5-best configurations are over 92% for all types, and those
[19]. Specifically, the ReLU activation function was applied against the test dataset are around 90%. This demonstrates that
before outputting features passed to the next layer in the graph the five types of graph convolution layers are effective in
convolutional layers. We adopted the average aggregation layer addressing this shooting scene classification problem. From
as the aggregation layer. The aggregated features were Table II, it is clear that the validation accuracies of the best
processed by the SoftMax function to estimate the probabilities configurations employing the three newly proposed types are
for the three categories of shooting scenes. We trained all the higher, surpassing 92.6%, than those of the conventional two.
neural networks (with any configuration and in each trial) over These facts indicate that their ability to accommodate negative
100 epochs using a mini-batch size of 1,024 and employed the edge weights have a beneficial impact on training. Even from
Adam optimizer with a step size parameter of 0.01. The the viewpoint of the average validation accuracies of the 5-best
negative log-likelihood function was adopted as the loss configurations, it is evident that the three newly proposed types
function. For each configuration of all the neural networks, we show higher performance compared to the conventional two.
selected the parameter set that achieved the highest accuracy in
the validation dataset across all epochs and all trials, as the best- On the other hand, Table III reveals that the conventional
trained parameter set of the configuration. The implementations transformer type [21], despite having the lowest validation
and settings were managed using Python 3 and the PyTorch accuracy among the five types, reaches the highest accuracy of

Authorized licensed use limited to: INSTITUTO FEDERAL DE SAO PAULO. Downloaded on May 07,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
the best configuration at 90.145%. However, the second-highest TABLE IV. TEST ACCURACIES (%) IN ABLATION STUDY
ON ADJACENCY INFORMATION
accuracy is recorded by the new type with the hyperbolic sine
function (i.e., (5)) at 90.112%, a marginal difference of only Ablation Scheme Average Acc. of 5-best Configs.
0.032 percentage points. Note that this type (5) also holds the Without Ablation (Original) 90.048
second-highest accuracy against the validation dataset. With Ablation Scheme (i) 89.622
Furthermore, the highest average accuracy of the 5-best With Ablation Scheme (ii) 89.909
configurations is observed with this type (5), which is 0.225
percentage points higher than that with the conventional differences, in particular, the difference from that with the
transformer type [21]. These suggest that this type (5) exhibits ablation scheme (i), i.e., the case of removing edges between all
considerably more stable performance compared to the nodes, were smaller than we initially expected. This may be
conventional transformer type [21]. because the shooting scene classification problem is too simple,
Comparing the three new types proposed in this study, the as similarly suggested in the previous study [19]. That is, the
new types with the simple linear function (6) and the hyperbolic shooting scene classification problem may not strongly require
tangent function (7) demonstrated nearly equivalent the adjacency information. Therefore, applications to more
performance to the new type with the hyperbolic sine function difficult problems might demonstrate clearer differences.
(5) on the validation dataset. However, contrary to our initial
expectations, they exhibited slightly inferior performance on the VI. CONCLUSION AND FUTURE WORK
test dataset. Notably, their average accuracies of the 5-best In this study, we proposed new graph convolutional layers
configurations are approximately 0.4 percentage points lower that are considered more effective for enhancing the formation
than that of the new type with the hyperbolic sine function (5). analysis method proposed in our previous study [19].
These suggest that these two new types (6) and (7) may have Furthermore, we validated the analysis method incorporating
lower stability. Moreover, it is validated that the performance the proposed layers by applying it to a soccer shooting scene
depends not only on the ability to accommodate negative edge classification problem. Additionally, we conducted an ablation
weights but on the function shape to compute the edge weights. study on the adjacency information input. Future work includes
applications to more challenging problems, aiming to reveal a
D. Ablation Study on Adjacency Information clearer distinction in the performance of different graph
In such a formation analysis method, it is an interesting convolutional layers.
matter to investigate how important the adjacency information
used as input is. Therefore, we also conducted an ablation study ACKNOWLEDGMENT
on the adjacency information after the validation described We would like to express our gratitude to DataStadium Inc.
above. However, in our investigation, we could not find a [38] for providing the data.
scheme for conducting an ablation study on the graph structure
of graph neural networks, applicable to our method. Therefore, REFERENCES
we considered two simple schemes: (i) removing edges between [1] T. Taki and J. Hasegawa, “Visualization of dominant region in team
all nodes, resulting in an edgeless graph where no nodes are games and its application to teamwork analysis,” In Proc. of Computer
connected to each other, and (ii) adding edges between all Graphics International 2000, pp. 227–235, Jun. 2000.
nodes, resulting in a complete graph where each node is directly [2] A. Fujimura and K. Sugihara, “Geometric analysis and quantitative
evaluation of sport teamwork,” Syst. Comput. Jpn., vol. 36, no. 6, pp. 49–
connected to every other node. It should be noted that the latter 58, Apr. 2005.
might not seem like a typical ablation scheme at first glance; [3] S. Fonseca, J. Milho, B. Travassos, and D. Araújo, “Spatial dynamics of
however, we consider this to also be an ablation scheme in the team sports exposed by Voronoi diagrams,” Hum. Mov. Sci., vol. 31, no.
sense of removing the influence from the input graph structure. 6, pp. 1652–1659, Dec. 2012.
In such an ablation study, we targeted our formation analysis [4] D. J. T. Sumpter, Soccermatics: Mathematical Adventures in the Beautiful
method enhanced by the new type of layer with the hyperbolic Game, 1st ed., London: Bloomsbury Sigma, May 2016.
sine function (5), since this shows the stable performance in the [5] S. Takahashi and M. Haseyama, “A note on network analysis based
detection of important player and similar scenes in soccer videos,” ITE
validation described above. Also for stability, we adopted the Tech. Rep., vol. 38, no. 51, pp. 1–4, Dec. 2014 (in Japanese).
average test accuracies of the 5-best configurations as the [6] S. Chawla, J. Estephan, J. Gudmundsson, and M. Horton, “Classification
performance measure. of passes in football matches using spatiotemporal data,” ACM Trans.
Spat. Algorithms Syst., vol. 3, no. 2, article no. 6, pp. 1–30, Aug. 2017.
Table IV lists the average test accuracy of the 5-best [7] T. Narizuka and Y. Yamazaki, “Clustering algorithm for formations in
configurations for each type of graph convolution layers. Here, football games,” Sci. Rep., vol. 9, article no. 13172, pp. 1–8, Sep. 2019.
the 5-best configurations for all cases were fixed as those [8] T. Mimura and Y. Nakada. “Quantification of pass plays based on
without any ablation. As demonstrated in this table, the average geometric features of formations in team sports,” In Proc. of the 10th
test accuracy of the original method is 0.426 percentage points International Symposium on Information and Communication
Technology, pp. 306–313, Dec. 2019.
higher than that with the ablation scheme (i) and 0.139
[9] T. Mimura and Y. Nakada. “Enhancements of pass play quantification
percentage points higher than that with the ablation scheme (ii). method with geometric features of formations,” In Proc. of 2022
This implies that the adjacency information used as input is International Conference on Engineering and Emerging Technologies, pp.
indeed effective in the formation analysis method. However, the 1–6, Oct. 2022.

Authorized licensed use limited to: INSTITUTO FEDERAL DE SAO PAULO. Downloaded on May 07,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.
[10] B. Ivanovic, E. Schmerling, K. Leung, and M. Pavone. “Generative [30] J. Albert, M. E. Glickman, T. B. Swartz, R. H. Koning, Handbook of
modeling of multimodal multi-human behavior,” In Proc. of 2018 Statistical Methods and Analyses in Sports, 1st ed., Boca Raton: CRC
IEEE/RSJ International Conference on Intelligent Robots and Systems, Press, 2017.
pp. 3088–3095, Oct. 2018. [31] D. Link, Data analytics in professional soccer : performance analysis
[11] R. A. Yeh, A. G. Schwing, J. Huang, and K. Murphy, “Diverse generation based on spatiotemporal tracking data, Wiesbaden: Springer Vieweg,
for multi-agent sports games,” In Proc. of 2019 IEEE Conference on 2018.
Computer Vision and Pattern Recognition, pp. 4610–4619, Jun. 2019. [32] T. A. Severini, Analytic Methods in Sports: Using Mathematics and
[12] H.-Y. Hsieh, C.-Y. Chen, Y.-S. Wang, and J.-H. Chuang, Statistics to Understand Data from Baseball Football Basketball and Other
“BasketballGAN: generating basketball play simulation through Sports, 2nd ed., Boca Raton: CRC Press, 2020.
sketching,” In Proc. of the 27th ACM International Conference on [33] D. Araújo, M. S. Couceiro, L. Seifert, H. Sarmento, and K. Davids,
Multimedia, pp. 720–728, Oct. 2019. Artificial Intelligence in Sport Performance Analysis, 1st ed., New York:
[13] M. Alcorn and A. Nguyen, “Baller2vec: a multi-entity transformer for Routledge, 2021.
multi-agent spatiotemporal modeling,” arXiv preprint arXiv:2102.03291, [34] A. Butterworth, Professional Practice in Sport Performance Analysis, 1st
pp. 1–14, Feb. 2021. ed., New York: Routledge, 2023.
[14] M. Alcorn and A. Nguyen, “Baller2vec++: a look-ahead multi-entity [35] J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, and M.
transformer for modeling coordinated agents,” arXiv preprint Sun, “Graph neural networks: A review of methods and applications,” AI
arXiv:2104.11980, pp. 1–13, Apr. 2021. Open, vol. 1, pp. 57–81, 2020.
[15] Y. Honda, R. Kawakami, R. Yoshihashi, K. Kato, and T. Naemura, “Pass [36] Z. Wu, S. Pan, F. Chen, G. Long, C. Zhang, and P. S. Yu, “A
receiver prediction in soccer using video and players’ trajectories,” In comprehensive survey on graph neural networks,” IEEE Trans. Neural
Proc. of 2022 IEEE/CVF Conference on Computer Vision and Pattern Netw. Learn. Syst., vol. 32, no. 1, pp. 4–24, Jan. 2021.
Recognition Workshops, pp. 3502–3511, Jun. 2022. [37] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
[16] P. Bauer, G. Anzer, and J. W. Smith, “Individual role classification for Ł. Kaiser, and I. Polosukhi, “Attention is all you need,” Advances in
players defending corners in football (soccer): Categorisation of the Neural Information Processing Systems, vol. 30, pp. 1-11, Jun. 2018.
defensive role for each player in a corner kick using positional data,” J. [38] DataStadium Inc.: [Link]
Quant. Anal. Sports, vol. 18, no. 2, pp. 147–160, Nov. 2022.
[39] PyG Team, PyG Documentation:
[17] T. Kaneko, R. Kawakami, T. Naemura, and N. Inoue, “Augmenting pass [Link]
prediction via imitation learning in soccer simulations,” In Proc. of 2024
IEEE/CVF Conference on Computer Vision and Pattern Recognition
Workshops, pp. 3194–3203, Jun. 2024.
[18] Z. Wang, P. Veličković, D. Hennes, N. Tomašev, L. Prince, M. Kaisers,
Y. Bachrach, R. Elie, L. K. Wenliang, F. Piccinini, W. Spearman, I.
Graham, J. Connor, Y. Yang, A. Recasens, M. Khan, N. Beauguerlange,
P. Sprechmann, P. Moreno, N. Heess, M. Bowling, D. Hassabis, and K.
Tuyls, “TacticAI: an AI assistant for football tactics,” Nat. Commun., vol.
15, no. 1, pp. 1–13, Mar. 2024.
[19] S. Takagi and Y. Nakada, “Formation analysis method for team sports
using deep graph convolutional neural network with geometric formation
features as input,” In Proc. of 2023 IEEE International Conference on
Multimedia Big Data, pp. 27-31, Dec. 2023.
[20] C. Morris, M. Ritzert, M. Fey, W. L. Hamilton, J. E. Lenssen, G. Rattan,
and M. Grohe, “Weisfeiler and leman go neural: higher-order graph neural
networks,” In Proc. of 2019 AAAI Conference on Artificial Intelligence,
vol. 33. no. 1. pp. 4602–4609, July 2019.
[21] Y. Shi, Z. Huang, S. Feng, H. Zhong, W. Wang, and Y. Sun, “Masked
label prediction: unified message passing model for semi-supervised
classification,” In Proc. of the 13h International Joint Conference on
Artificial Intelligence, pp. 1548–1554, Aug. 2021.
[22] J. Gudmundsson and T. Wolle, “Football Analysis Using Spatio-Temporal
Tools,” Comput. Environ. Urban Syst., vol. 47, pp. 16–27, Sep. 2014.
[23] J. Gudmundsson and M. Horton, “Spatio-temporal analysis of team
sports,” ACM Comput. Surv., vol. 50, no. 2, pp. 1–34, Mar. 2018.
[24] F. Lorda, D. B. Pyne, M. Welvaert, and J. K. Mara, “Methods of
performance analysis in team invasion sports: a systematic review,” J.
Sports Sci., vol. 38, no. 20, pp. 2338–2349, Jun. 2020.
[25] T. Blobel, M. Rumo and M. Lames, "Sports information systems: a
systematic review", Int. J. Comput. Sci. Sport, vol. 20, no. 1, pp. 1-22,
May 2021.
[26] Z. Pu, Y. Pan, S. Wang, B. Liu, M. Chen, H. Ma, and Y. Cui, “Orientation
and decision-making for soccer based on sports analytics and AI: a
systematic review," IEEE/CAA J. Autom. Sin., vol. 11, no. 1, pp. 37-57,
Jan. 2024.
[27] I. Fister, I. Fister Jr., and D. Fister, Computational Intelligence in Sports,
1st ed., Berlin: Springer, 2015.
[28] D. J. T. Sumpter, Soccermatics: Mathematical Adventures in the Beautiful
Game, London: Bloomsbury Sigma, 2016.
[29] A. Jayal, A. McRobert, G. Oatley and P. O'Donoghue, Sports Analytics:
Analysis Visualisation and Decision Making in Sports Performance, New
York: Routledge, 2017.

Authorized licensed use limited to: INSTITUTO FEDERAL DE SAO PAULO. Downloaded on May 07,2025 at [Link] UTC from IEEE Xplore. Restrictions apply.