5178 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 29, NO.

12, DECEMBER 2023

Team-Builder: Toward More Effective

Lineup Selection in Soccer
Anqi Cao , Ji Lan , Xiao Xie , Hongyu Chen, Xiaolong Zhang, Hui Zhang , and Yingcai Wu

Abstract—Lineup selection is an essential and important task in soccer matches. To win a match, coaches must consider various
factors and select appropriate players for a planned formation. Computation-based tools have been proposed to help coaches on this
complex task, but they are usually based on over-simplified models on player performances, do not support interactive analysis, and
overlook the inputs by coaches. In this article, we propose a method for visual analytics of soccer lineup selection by tackling two
challenges: characterizing essential factors involved in generating optimal lineup, and supporting coach-driven visual analytics of lineup
selection. We develop a lineup selection model that integrates such important factors, such as spatial regions of player actions and
defensive interactions with opponent players. A visualization system, Team-Builder, is developed to help coaches control the process of
lineup generation, explanation, and comparison through multiple coordinated views. The usefulness and effectiveness of our system
are demonstrated by two case studies on a real-world soccer event dataset.

Index Terms—Sports visualization, lineup selection, design study

1 INTRODUCTION Specifically, different players have various playing styles and

perform differently on the aspect of offense and defense. The
soccer matches, team lineup selection is vital and a
I N
proper lineup with effective tactics can significantly
increase the chance of winning [1]. Numerous examples
coach has to choose a lineup carefully by considering the syn-
ergy between team tactics and each player’s offensive and
defensive skills. This process usually involves extensive data
among top leagues and national teams illustrate the impor-
analytical activities.
tance of lineup selection. For instance, the unexpected suc-
Recently, various computational tools have been pro-
cess by the team Leicester City in the 2015/16 season of the
posed to help coaches find suitable lineups [5]. However,
English Premier League is partially attributed to excellent
those approaches mainly use complicated models to gener-
lineup selections that lead to its effective counter-attack tactic
ate the best lineup automatically, and often overlook some
[2]. In recently held EURO 2020, a number of national teams,
important information in lineup selection, such as team tac-
such as Denmark, Switzerland, and Czech, selected suitable
tics. Without the knowledge of the computational processes
lineups and adopted effective tactics, which helped them
involved in the models, coaches often find it a big challenge
defeat their opponent teams that included star players [3].
to understand how the results are obtained; and unable to
Selecting a good lineup is not a trivial task. Coaches need to
incorporate more diverse information, these models cannot
first choose team tactics and then choose players under vari-
support more in-depth analysis of lineups. Therefore, such
ous criteria and through multiple comparisons among the
automatic lineup selection models are seldom adopted by
skills and performances of players [4]. Thus, lineup selection
coaches in professional soccer teams [5].
can be described as a problem that requires trade-offs between
In addition, lineup selection often needs interactive analy-
multiple criteria, some of which could be contradictory.
sis to try and compare different options and parameters
based on real-time information (e.g., possible strategies of

Anqi Cao, Ji Lan, Hongyu Chen, and Yingcai Wu are with the State Key the opponent team). Existing methods for lineup analysis are
Lab of CAD&CG, Zhejiang University, Hangzhou 310027, China. E-
mail: {caoanqi, lanjizju, chenhy7820, ycwu}@zju.edu.cn.
usually weak in providing interactive tools. Although vari-

Xiao Xie and Hui Zhang are with the Department of Sports Science, Zhe- ous interactive visualization systems have been available for
jiang University, Hangzhou 310027, China. E-mail: {xxie, zhang_hui} soccer data analysis, such as positions and actions [6], forma-
@zju.edu.cn. tion variation [7], passing patterns [8], and migration [9],

Xiaolong Zhang is with the College of Information Sciences and Technol-
ogy, Pennsylvania State University, State College, PA 16801 USA. tools for interactive lineup analysis and selection are rare.
E-mail: [email protected]. In this research, we propose a visualization system to pro-
Manuscript received 23 December 2021; revised 23 August 2022; accepted 1 vide a systematic solution for the lineup selection problem. The
September 2022. Date of publication 16 September 2022; date of current ver- system can support the exploration and evaluation of various
sion 10 November 2023. lineup options recommended by the model. Our efforts are
The work was supported in part by NSFC under Grant 62072400, in part by the
Collaborative Innovation Center of Artificial Intelligence by MOE and Zhejiang made to address two main challenges. The first challenge is
Provincial Government (ZJU), and in part by Zhejiang Lab (2021KE0AC02). related to the difficulty in characterizing and modeling the
(Corresponding author: Xiao Xie.) complex factors that should be considered in lineup selection.
Recommended for acceptance by J. Fekete. In soccer matches, lineup selection is influenced by not only the
This article has supplementary downloadable material available at https://doi.
org/10.1109/TVCG.2022.3207147, provided by the authors. performances of individual players but also the teamwork
Digital Object Identifier no. 10.1109/TVCG.2022.3207147 among players [10], [11], [12]. Previous studies have modeled
1077-2626 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See ht_tps://www.ieee.org/publications/rights/index.html for more information.
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
CAO ET AL.: TEAM-BUILDER: TOWARD MORE EFFECTIVE LINEUP SELECTION IN SOCCER 5179

the teamwork in basketball [13], [14] and soccer [15], [16] to data of events during a match, data has been collected through
help coaches select the best lineup. However, other important videos directly, and movement trajectories of players and the
and heterogeneous factors of teamwork, such as the interac- ball. Soccer event data describe the spatio-temporal events that
tions with the opponent and corresponding tactics preferred occurred in matches. Such kind of data is widely used for pass-
by coaches, have not been considered yet [17]. It is non-trivial ing pattern analysis and player performance analysis [18]. To
to integrate those complex factors into existing models. help the investigation of passing patterns among players, one
Another challenge lies in the difficulty in designing a system- of the most essential aspects of soccer event data analysis,
atic tool to assist soccer coaches to make the best lineup deci- researchers have developed systems like SoccerStories [38], a
sion. As the experiences and preferences of soccer coaches vary visualization system on player actions and spatial passing pat-
significantly, the factors that should be considered in the lineup terns, and PassVizor [8], a visual analytical system for dynamic
selection model could be significant, so interactive control over passing patterns. Malqui et al. [39] also developed a visual ana-
these factors and relevant parameters is required. Coaches also lytics system to discover soccer passing strategies based on
need to evaluate different aspects of a selected player such as flow motifs. Compared with event data, soccer match videos
individual features and interactions with other players in and trajectory data are more detailed and include the positions
matches, multiple criteria of lineups such as the total number of all players and the ball during the whole match. These data
of passing, shot, and interception, and the predicted match could support diverse analyses such as video analysis, trajec-
results of each lineup. The system should also allow visual tory analysis, and formation analysis [18]. To facilitate soccer
comparison of players and lineups recommended by the model match video analysis, Stein et al. [40], [41] proposed an auto-
according to various criteria. matic video annotation technique that can integrate visualiza-
Teamed up with the domain experts, we develop a new tion of player movements with videos. In addition to some
model for lineup selection and design a system, Team- methods for the computation of player movement trajectories,
Builder, based on the model. Assisted by soccer coaches and such as a trajectory search method proposed by Shao et al. [42],
soccer data analysts, we synthesize and operationalize vari- and trajectory aggregation methods by Sacha et al. [43], various
ous factors for lineup selection (e.g., tactical preferences and visualization-based systems have been developed to support
opponent information), and enhance a teamwork-based interactive analytics of player movement trajectories.
model [15] to integrate these identified factors in the model. Andrienko et al. used player trajectory data to show player
Team-Builder, a visual analytics system, helps coaches adjust defense [44] and team tactics [6]; Janetzko et al. [45] designed a
the model according to their preferences and evaluate model system to analyze player performances based on extracted
results. Our research makes the following contributions: player trajectories; Machado et al. [46] provided a heatmap-
based visualization for team formation identification; and For

we formalize the characterization of the problem of soc- Vizor [7] developed a tailored flow-based visualization to rep-
cer lineup selection, including identification and resent the change of team formations and a system for a system-
modeling of key factors, specification of preferences, atic analysis of the formations. Besides, visual analytics systems
and evaluation of player and lineup candidates; based on soccer trajectory data also contain other aspects such

we synthesize some key factors for lineup selection as what-if analysis for player passing decisions [47] and perfor-
and developed a teamwork-based model to integrate mance analysis based on player coordination [48].
these factors; and Although these studies can help soccer data analysis one

we develop a visual analytics system to support explo- way or another, few could support visual analytics for soccer
ration and evaluation of lineup recommendations by lineup selection, because of the difference in analysis tasks.
the model.

2 RELATED WORK 2.2 Soccer Lineup Selection

Soccer lineup analysis has attracted extensive research atten-
In this section, we review research on soccer data visualiza-
tion, especially on the aspects of formation detection [49], [50]
tion and soccer lineup selection methods.
and evaluation [51]. In particular, lineup selection is vital in soc-
cer matches, and an increasing number of studies in this field
2.1 Visualization of Soccer Data have emerged during recent years [5]. Some research has pro-
The interests of analyzing sports data with visualization vided models to find the best soccer lineup by aggregating the
techniques have increased significantly [18], [19], as seen in performances of individual players. Boon and Sierksma [1] pio-
work on basketball [20], [21], [22], baseball [23], [24], [25], neered a study to solve the problem of selecting an optimal soc-
[26], racquet sports [27], [28], [29], [30], and other events cer lineup. They estimated the weights of individual player
[31], [32], [33]. There was considerable visualization work to performances on different positions, and selected the optimal
analyze soccer match data from diverse aspects. For exam- lineup by maximizing the total performance value. Tavana
ple, Perin et al. developed A table! [34] and Gap Charts [35] et al. [52] further defined soccer lineup selection as a multi-crite-
to visualize the temporal evolution of soccer team rankings; ria decision-making (MCDM) problem, and let the important
and Rusu et al. [36], [37] developed metaphor-based visuali- weights of player performance attributes be controlled by users,
zation to compare performances among different players such as coaches who have better knowledge about players and
with statistical indicator analysis. positions. Based on the definition of the lineup selection prob-
Analytical methods in soccer data visualization systems lem, they proposed a fuzzy inference system and a linear pro-
have become more and more sophisticated. Involved data is no gramming model to maximize the overall lineup score.
longer limited to simple statistics and includes fine-grained Similarly, Ozceylan [4] also modeled soccer lineup selection as
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
5180 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 29, NO. 12, DECEMBER 2023

an MCDM problem, and maximized the sum of player value TABLE 1

based on player property weights under multiple performance The Event Attributes
criteria. In addition to MCDM-based methods, other appr-
Attribute Description
oaches were also proposed for selecting soccer lineups accord-
ing to player individual performances. For instance, Merig
o Event Type A technique that a player takes to deal with
and Gil-Lafuente [53] presented a soccer player selection model the ball (pass, cross, free-kick, dribble, shot,
foul, tackle, interception, offside, etc.).
based on business decision-making methods, and Al-Shboul Event Result Result of the event (succeeded or failed).
et al. [54] utilized neural networks to decide the best lineup Time Time as the event occurred in a match.
with the highest winning rate. Positions The starting and ending coordinates of the
Soccer is a team sport, so teamwork is an essential aspect event in the field.
when coaches select players [10], [11], [12]. However, this Player The player who started the event.
team factor was largely overlooked by the methods dis- Player Role The role of the player (goalkeeper, defender,
midfielder, or forward).
cussed above. To integrate this important factor in lineup
Team The team of the player.
selection, Beal et al. [15] constructed a framework to quan- Match ID The index of the match.
tify direct interactions among players from the same team
with network-based metrics and found the best team using
the mixed-integer programming method. Bransen et al. [16]
improved this framework with a data-driven approach to A tactic refers to a set of similar actions or phases that
calculate interactions by the change of goal probabilities. leads to the same tactical effects. We use the definition of an
However, these models do not consider factors related to offensive tactic as a cluster of phases in similar spatial
the opponent and the corresponding tactical inputs from regions with a purpose to drive the ball to the goal of the
coaches. The lack of such factors limits their applications, opponent [57]. In soccer event analysis, defense could be
because lineup selection can vary from opponent to oppo- reflected from defensive actions to intercept the passing of
nent and different coaches may have different opinions the opponents or take the ball directly. Therefore, we use
when competing with the same team [52]. the definition of a defensive tactic as a set of actions in simi-
To address the limitations and challenges in the visual lar spatial regions to defend the goal [17].
analytics of soccer lineup selection, we propose a team- An offensive interaction based on soccer events data is com-
work-based method that considers the performances of posed of two consecutive actions by two players in the same
individual players, information on the opponent team, and phase [15]. It is the basic unit to evaluate how players cooper-
the inputs from the coach. ate through the movement of the soccer ball. For instance, if a
player is surrounded by several opponents and has difficulty
in taking further actions, the player will pass the ball to a
3 BACKGROUND AND SYSTEM OVERVIEW teammate in a better position. If an offensive interaction is
In this section, we first introduce the relevant terminology successful, the chance to score a goal can be increased. There-
in soccer analysis and the data used in this research. Then, fore, offensive interactions between players are an essential
we describe our work to develop the design requirements indicator in soccer lineup selection [15], [16].
assisted by domain experts. A defensive interaction could be defined as two consecu-
tive actions of players from different teams. Specifically, a
defensive interaction is composed of the last action of the
3.1 Background
previous phase and the first action of the next phase. In soc-
Soccer is a highly dynamic team sport. In a match, two teams
cer event data, the confrontation between two teams can be
attempt to shoot the soccer ball into the goal of the opponent,
reflected by the soccer ball transiting through defensive
and the team with more goals wins the match. Each team has
actions [17]. Similar to offensive interaction, a successful
eleven players on the field, and they assume one of the four
defensive interaction will increase the goal chance of a
roles: goalkeeper, defender, midfielder, and forward.
team, and also decrease the goal chance of the opponent.
An action is an event started by a player who has the ball
A lineup is the eleven players of a team selected by the
under control or tries to take over the control of the ball. It is a
coach to appear in a match. Before a match, the coach and
fundamental element in soccer event analysis [55]. The details
staff members will analyze the offense and defense of the
of the event attributes are shown in Table 1 [55], [56]. The
opponent team, and select well-performed players with the
actions could be further divided into offensive and defensive
consideration of such opponent information. Considering
actions according to the event type. We refer the actions with
the importance of teamwork, the lineup also considers how
event types of foul, tackle, interception to defensive actions and
those with other types to offensive actions. Generally, the offen- well those selected players can work together.
sive actions indicate how the players take or pass the ball to cre-
ate a goal, and the defensive actions aim to take back the
possession of the ball from the opponent team. 3.2 Data Description
A phase is a series of consecutive actions performed by The data used in our analysis is from the soccer event dataset
players in the same team with the ball under control [57]. A published by Pappalardo et al. [56]. The dataset contains not
phase starts when a player takes control of the ball, and only spatio-temporal event attributes, but also contextual
ends when the ball is lost to the opponent. During a phase, data, such as players involved in each event, player team, and
the team controlling the ball tries to win a goal by successive match information. The primary event attributes involved in
actions, such as passing, dribbling, and shooting. the analysis are shown in Table 1.
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
CAO ET AL.: TEAM-BUILDER: TOWARD MORE EFFECTIVE LINEUP SELECTION IN SOCCER 5181

The data we used for analysis contains all matches in and the prototype. After several rounds of iterations,
the five major European soccer leagues in the season design requirements and the prototype were stabilized. The
2017/18. In detail, the data consists of 380 matches from timeline of the design iteration process is provided in our
Spanish La Liga, 380 matches from English Premier League, supplementary material, available online.
380 matches from Italian Serie A, 306 matches from Ger- Design Requirement Development. Our data indicate that a
man Bundesliga, and 380 matches from French Ligue 1. system to support lineup selection need to provide tools for
The total number of matches is 1,224, including 98 teams tasks at three levels: tactic-level exploration, player-level inves-
and 4,229 players. tigation, and lineup-level comparison.
Tools for tactic-level exploration aim to support coaches to
identify tactical preferences when facing a given opponent
3.3 Requirement Analysis by integrating their domain knowledge of lineup selection.
We worked with a team of three domain experts to learn
about the methods for lineup selection. The expert team T1 Supporting the exploration of different categories of tac-
included a coach of a professional soccer team, a professor tics. When selecting a lineup before a match, coaches
of physical education who is a senior sports analyst with need to obtain an overview of tactic categories of
decades of experience, and a doctoral student in physical their team and those of the opponent. Providing nec-
education who was a professional soccer player in a top essary tactic information can help coaches compre-
national-league team. hend the characteristics of both teams and further
Our goals are multi-fold. First, we wanted to characterize decide the tactics for a match.
the problem of soccer lineup selection through meetings T2 Supporting the exploration of available tactics and their
and interviews with the experts. Second, we expected to effectiveness according to spatial regions when facing a
develop a lineup selection model based on what we have given opponent. Coaches tend to analyze the tactics of
learned from the experts. Finally, we sought help from the opponent and prepare corresponding tactics on
them in developing design ideas and design requirements the aspects of offensive and defensive. For instance,
for our visual analytics system. if the opponent team tends to attack from the left
Characterization of the Problem. We reviewed relevant lit- side, coaches will choose players who are good at
erature [1], [4], [15], [16], [52], [53], [54] to summarize the defending on the right. This information can provide
current problem characterization of soccer lineup selec- references for the selection of players who are suit-
tion and held a series of meetings with the experts. Dur- able for attacking the weakness and defending the
ing the meetings, we mainly discussed two questions: strike of the opponent team.
whether the problem characterization in literature is con- Tools for player-level investigation can help understand the
sistent with their experiences, and how lineup selection is player selection process and interpret the selection results.
made in real soccer matches. We also had interviews with
each expert to learn about their own preference and crite- P1 Supporting the comparison of player performances under
ria used in lineup selection. We recorded the discussion multiple criteria. After deciding the tactics for the
process in the meetings and the answers of each expert in match, coaches would like to select several players
the interviews. Based on data collected from the meeting who performed well within the selected tactics as
discussions and interviews, we drafted the initial require- the core players in the lineup. They need to compare
ments for system design. the performances of these players under different
Development of the Lineup Selection Model. We applied a criteria. Thus, it is necessary to support efficient
series of models and discussed if the model results meet comparison among players to help coaches select
their expectations and how to improve the models with the suitable core players.
experts. They compared these models, and stated that team- P2 Supporting the explanation of player selection in the lineup
work-based models are more practical and straightforward result. The lineup selection model can recommend
than deep learning models, because the results of the former players based on the coaches’ preferences. Coaches
match their domain experience better. Furthermore, the require to know why a player is selected by the model
experts suggested that the current teamwork-based models to determine whether it is reasonable to include the
could be improved by considering other essential informa- player in real matches. The lineup selection result is
tion such as the opponent and tactical choices of coaches. associated with players’ individual performances
Therefore, we constructed a teamwork-based model by inte- and interactions with other players. The tool needs to
grating opponent and tactical information. provide these details.
Iteration of Visual Design. After the model development, we Tools for lineup-level comparison assist to obtain information
designed a visual design prototype based on the initial require- of multiple results and to decide the most suitable lineup.
ments and our model. The prototype included tactic detection
and automatic lineup recommendation according to preferred L1 Supporting the comparison of expected match results and
tactics of the users. We presented the prototype to the experts reliability of different lineups. Coaches may generate
for feedback. They commented that explaining why a player is several candidate lineups under different preferen-
selected by the model is also demanded by coaches. Besides, ces and select one of them considering both expected
comparison among multiple lineup selection results and what- match result and reliability. The expected match
if analysis would help them decide the most suitable lineup. result is the probability of winning/drawing/losing
Their suggestions helped us improve design requirements the match, and the reliability means the appearance
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
5182 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 29, NO. 12, DECEMBER 2023

fp1 ; . . . ; pnP g (n indicates the amount of the elements in set

) and a lineup as one of the subsets of P with eleven
elements.
Soccer lineup selection is a complex process. We simplify
the process and only consider the most essential factors in
lineup selection. According to previous work [1], we define
the most suitable lineup as the one with the best team per-
formance. Based on the definition, the lineup selection pro-
cess can be modeled as an optimization problem of
choosing eleven players with the maximum value of a spe-
cific team performance metric [1]. In soccer matches, the
team performance can be estimated by the historical match
data and the physical and mental conditions of players. We
further simplify the process by estimating the team perfor-
mance only using the historical match data.
Several metrics have been proposed to select the optimal
Fig. 1. Major components of the system and their relationships. The sys-
tem has three components: the data processing component, the player lineup, including the match winning rate [54], the total per-
selection model component, and the visualization component. formance score [1], [4], [52], and the teamwork value [15],
[16]. During the cooperation with our experts, we found
time of players in previous matches. Tools should be that the teamwork value is more practical than other metrics
provided to help compare these metrics and decide because interaction among the players is also important for
on the most appropriate lineup. a match. Therefore, we use teamwork value as the team per-
L2 Supporting the comparison of predicted statistical indica- formance metric that is maximized in our model.
tors of different lineups. When the lineup used for a The teamwork value among players refers to the composi-
match is decided, coaches demand to obtain the pre- tion of the total individual performance value and the total
dicted statistical indicators of the lineup and com- interaction value of all players in the lineup. In soccer matches,
pare them with those of other candidate lineups. the interactions among the players could be divided into direct
These indicators could help coaches evaluate interactions completed through ball transition between two
whether the tactical characteristics of the lineup players, and indirect interactions accomplished by the coordi-
meet their expectations. Such detailed knowledge nated movements of the players [15]. We only consider the
could also facilitate the choice of suitable match direct interactions between two players because they are the
strategies such as build-up and counter-attack. most basic and important parts of player teamwork [15]. Thus,
in the calculation of the teamwork value, a soccer match can be
3.4 System Overview treated as a sequence of player actions A ¼ fap1 ; . . . ; apnA g,
We develop Team-Builder, a web-based application to sup- where p 2 PM is the player who performs the actions and PM
port visual analytics of soccer lineup selection. The system is the set of all players involved in the match (blue & orange
has three components: a data processing component, a player objects in Fig. 2).
selection model component, and a visualization component Some teamwork-based models have been proposed for
(Fig. 1). The data processing component supports the extrac- soccer lineup selection [15], [16]. However, those models
tion of phases from the raw event dataset and the calculation have some critical limitations in real-world practices.
of the value for each action and interaction. The player selec- Neglecting Spatio-Temporal Information. The original meas-
tion model component establishes the teamwork-based urements of player interaction Iðpi ; pj Þ considered all the
p pj
model for the selected target team with the data from the consecutive two actions ðaki ; akþ1 Þ of a certain pair of play-
data processing component. The visualization component ers ðpi ; pj Þ in the same team and aggregated the values
includes the user interface of the system for interactive lineup directly (black objects in Fig. 2). However, these measure-
selection and consists of three views: the tactic view, the ments do not reflect the tactical preference of coaches (e.g.,
player view, and the lineup view. The system was imple- wing-attack, mid-attack) because the spatio-temporal infor-
mented with MongoDB for the data processing component, mation of phases is lost. Valuing teamwork without spatio-
Flask in Python for the player selection model component, temporal information is inadequate.
and React for the visualization component. Ignoring the Opponent of the Team. Player interactions with
the opponent indicate defensive features between players
4 TEAMWORK-BASED MODEL FOR PLAYER from different teams, and are also an important factor when
coaches select the lineup. However, previous models did not
SELECTION
distinguish teamwork values with opponents and provided
In this section, we describe our teamwork-based model to the same solution for diverse opponents. Therefore, it is insuf-
generate lineups for soccer matches. ficient to generate solutions in practical matches.
To address these two limitations, we develop a new
4.1 Task Definition lineup selection model that considers the spatio-temporal
The major task of soccer lineup selection is to identify information and the opponent. As mentioned before, our
eleven players for a certain match from all players in the model only considers part of the problem and is just a step
team. We define the set of candidate players as P ¼ to understand the problem better. For the first limitation,
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
CAO ET AL.: TEAM-BUILDER: TOWARD MORE EFFECTIVE LINEUP SELECTION IN SOCCER 5183

Fig. 2. The structure of a soccer match. ai refers to an action, IO an

offensive interaction, and ID a defensive interaction. The gray objects
illustrate that in a soccer match, a new phase would start when the ball
possession has changed.
Fig. 3. The calculation of the teamwork value of Player p. pi refers to a
player in the same team and qi a player in the opponent team. IOk and
we integrate a tactic detecting model into the calculation of k
ID represent offensive and defensive interaction value between two
teamwork value to identify important interactions. The tac- players.
tic detecting model could find different tactics T ¼
ft1 ; . . . ; tnT g by spatio-temporal similarities. Coaches can As mentioned before, we only take the direct interactions
choose preferred tactics and the weights of corresponding into the calculation of the teamwork value. The teamwork
interactions will be increased in the calculation of the team- value of a certain player p is shown in Fig. 3, composed of
work value. With such a method, players who perform well the individual performance value, the offensive interaction
in the selected tactics will be more likely to be chosen. To values with the teammates selected in the lineup (Fig. 3 A),
deal with the second limitation, we divide the interactions and the defensive interaction values with the opponents
among players into offensive interactions IO and defensive (Fig. 3 B). Here we introduce the definition of those features.
interactions ID (black objects in Fig. 2). In this way, the For simplicity, we use P ¼ fp1 ; . . . ; pnP g to represent the set
teamwork value is associated with the actions of the of all candidate players in a team.
opponent. Individual Performance Evaluation. For a player pi 2 P , we
use the individual performance value VI ðpi Þ to indicate how
4.2 Tactic Detection well the player performed from the aspect of personal abil-
When selecting a soccer lineup, coaches usually pay atten- ity. According to existing research, individual performance
tion to the tactics used by the target team and its opponent can be measured by abstract statistic indicators such as
to evaluate the strengths and weaknesses of the two teams. passing network attributes [15] and the contribution to goal
Such information could provide vital references for coaches probability [55], [58]. In our case, we choose the approach
to select appropriate players to confront the opponent. that is based on the changes of goal probabilities for the
Thus, we utilize tactic detection models to extract offensive player’s own and opponent teams [55]. We calculate the
tactics [57] and defensive tactics [17] used by a specific performance value of each action on reward and risk. The
team. The detailed detection process of offensive and defen- calculation process is provided in our supplementary mate-
sive tactics is provided in our supplementary material, rial, available online. The individual performance value of a
available online. In our lineup selection model, coaches player is calculated by aggregating the performance value
could select the offensive and defensive tactics they desire of all actions of the player.
to use, and the value of interactions included in the chosen Offensive Interaction Value. For a pair of players ðpi ; pj Þ,
tactics would be allocated with an elevated weight. where pi ; pj 2 P , we use the offensive interaction value
IO ðpi ; pj Þ to indicate the effectiveness of player interactions
4.3 Teamwork Value Evaluation in offense (Fig. 3 A). Generally, coaches treat each phase as a
Both the individual player performance and the interactions complete offense process and the basic unit for team offense
among players are essential to lineup selection. Compared analysis [57]. Therefore, the interactions among players in
with the interactions among players in the same team, the the same team during the same phase could be regarded as
interactions between two teams indicate how the control of offensive interactions (black objects in Fig. 2). On the basis of
the ball transits, and can reflect defensive tactics. Thus, we the previous works, IO ðpi ; pj Þ is the aggregation of all interac-
decide to take these features to measure the teamwork value tions between pi and pj . The equation is as follows:
of a lineup. Based on the previous work [15], [16], we con-
X
n
90
struct a model for teamwork value calculation as follows: IO ðpi ; pj Þ ¼ ðwk  IOk ðpi ; pj ÞÞ  ; (2)
k¼1
Mðpi ; pj Þ

V ¼ 1  VI þ 2  IO þ 3  ID ; (1) where k refers to the index of the current interaction in the

set of all interactions between pi and pj , n represents the
where 1 þ 2 þ 3 ¼ 1. VI denotes the total individual per- total number of interactions between pi and pj , IOk ðpi ; pj Þ
formance value, IO and ID refer to the total offensive inter- means the offensive interaction value on the index k,
action value and the total defensive interaction value, Mðpi ; pj Þ is the number of minutes that pi and pj have
respectively. 1 , 2 , and 3 are the weights of different fea- played together, and wk is the weight of the corresponding
tures, which are estimated from historical match data. interaction. IOk ðpi ; pj Þ is gained by the addition of the values
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
5184 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 29, NO. 12, DECEMBER 2023

of the two consecutive actions involved in the interaction.

The calculation of the action value is the same as that for the
individual performance value. If a tactic is selected, all inter-
actions that belong to the tactic will have an elevated weight
wk > 1. The value is normalized by 90 minutes, the regular
time of a soccer match, to avoid ignoring players who sel-
dom played together but interacted effectively.
Defensive Interaction Value. For a pair of players from dif-
ferent teams ðpi ; qj Þ, where pi 2 P , qj 2 Q (Q is the set of
players of another team), we use the defensive interaction
value ID ðpi ; qj Þ to evaluate player defense (Fig. 3 B). The
beginning of each phase implies a successful defense. Thus,
a defensive interaction can be described by the last action in
the former phase and the first action in the latter phase Fig. 4. The evaluation results in the first experiment. The red parts repre-
sent the previous model and the green parts represent our model.
(black objects in Fig. 2). To gain a comprehensive defensive
interaction feature, we calculate the defensive interaction
value with all historical match data of the target team. To players who are both in the generated lineup and the real
cope with the specific opponent team in the match, we lineup to evaluate the models. Fig. 4 shows the comparison
assign an elevated weight to the interactions with the same of two models in terms of the averages of the same players
opponent in previous matches to indicate the importance of in model-generated lineups and real. As the result shows
those interactions. Similar to the offensive interaction value, that the lineups generated by our model (green objects in
the equation of defensive interaction is defined as follows: Fig. 4) are closer to real lineups chosen by coaches than
those by the previous model (red objects in Fig. 4). This
X
n
90 implies that when selecting the lineup, coaches consider not
k
ID ðpi ; qj Þ ¼ ðwk  ID ðpi ; qj ÞÞ  ; (3) only the individual performance and interactions with
k¼1
Mðpi ; qj Þ
teammates, but also interactions with opponent players.
Thus, considering defensive interactions in the calculation
k
where ID ðpi ; qj Þ indicates the defensive interaction value of the teamwork value can help coaches select better lineups
between pi and qj on the index k. The calculation of defen- in real matches.
sive interaction value is the same as offensive interaction In the second experiment, we evaluated that teamwork
value. value is an effective metric to measure team performance.
Please refer to our supplementary material for details, avail-
4.4 Solving the Optimization Equation able online.
We solve the optimal lineup by maximizing the teamwork
value defined in Equation (1) through integer linear pro-
5 VISUAL DESIGN
gramming. Specifically, we construct the optimization equa-
tion based on the work of Beal et al. [15] and Bransen et al. In this section, we describe the design of visualization and
[16]. We simplify the optimization process by only consider- interaction tools in our system based on the previous
ing the most important constraints, including the total requirement analysis.
player number, the team formation, and the players manu-
ally included by coaches and who cannot appear in the 5.1 Overview of Visual Design and User Interface
match. Please refer to our supplementary material for According to the requirements, we design a tactic view for
details, available online. tactic exploration (T1, T2), a player view for lineup genera-
tion and player investigation (P1, P2), and a lineup view for
4.5 Model Evaluation solution comparison (L1, L2).
Our model was evaluated with the soccer event dataset The workflow of the system is as follows. In the tactic
used in our analysis, including all matches in the 2017/18 view, users could select the target team and its opponent,
season from the five top European leagues [56]. In the and obtain an overview of all tactics used by both teams to
experiments, we split the matches in the dataset into two specify essential tactic categories (T1) (Figs. 5 A1 and 5 A2).
half-seasons for training and testing. Besides, we evaluated Then, users could explore confrontation tactics in the
the performance of our model on the matches from the dif- selected categories through ranking and filtering tools to
ferent five leagues. choose the desired tactic combination (T2) (Figs. 5 A3 and 5
In the first experiment, we evaluated our model with the A4). The player view provides a candidate player list to
same players in lineups in real matches, and compared the enable users to compare different players on the individual
results with the previous model that does not consider the performance and interactions with other players (P1) (Fig. 5
defensive interaction value [16]. We measured both models B2). Based on the comparison, users can include or exclude
by comparing how many players in each lineup generated certain players in the lineup. After the tactics and players
by a model match the players in real lineup. Specifically, we are set, users could acquire the generated lineup and further
trained the two models with the match data in the first half- adjust it in the lineup edit board of the player view (Fig. 5
season, and generated lineups with the two models for each B1). The player view also contains an explanation compo-
match in the second half-season. We counted the number of nent to explain why the players are selected in the
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
CAO ET AL.: TEAM-BUILDER: TOWARD MORE EFFECTIVE LINEUP SELECTION IN SOCCER 5185

Fig. 5. System user interface. The interface consists of three views: a tactic view (A), a player view (B), and a lineup view (C). The tactic view provides
confrontation tactic lists (A1, A2, A3, A4) to navigate tactics used in the lineup. The player view contains a lineup edit board (B1) for lineup generation,
a candidate player list (B2) for player constraint identification, and an explanation component for comprehension of the reason of the selection of a
player. The lineup view includes a candidate lineup list (C1) and lineup thumbnails (C2) for comparing multiple lineups.

generated lineup (P2) (Fig. 7). To identify the best lineup 5.2 Tactic View
under multiple criteria, users can record lineups and com- The tactic view is composed of a confrontation tactic list to
pare them in the lineup view (L1, L2) (Fig. 5 C). To assist the present all tactics used by the target team (Figs. 5 A2 and 5
analysis, we also embed customized diagrams (Fig. 6) in the A4) and its opponent (Figs. 5 A1 and 5 A3). The tactics are
system to encode team tactics. illustrated by tactic diagrams with tactic category and spa-
We use blue and orange color to code the target team and tio-temporal context (Fig. 6). In this view, users can directly
its opponent team, and purple and yellow color to code the select the target team and its opponent with the drop lists,
offensive and defensive interaction values through the filter tactics with essential tactic categories (T1), and navi-
whole user interface. gate the tactics to be applied for the match through ranking
by their usage rate and success rate (T2).
Tactic Diagrams. To facilitate exploration and comparison
among tactics (T2), we design tactic diagrams for offensive
and defensive tactics in soccer matches, respectively (Fig. 6).

Pitch division. When describing soccer tactics, the

pitch could be divided into nine spatial regions,
including the defensive third, the middle third, and
the attacking third [17]. Therefore, we use this divi-
sion method to simplify the tactic representation in
our design (Fig. 6 B).

Offensive tactic. An offensive tactic is a cluster of
phases on similar spatial regions. To visualize an
offensive tactic, we first aggregate all phases in the
tactic to an abstracted sequence that can represent
the spatio-temporal feature of the tactic [43] (Fig. 6
A). Then, based on the pitch division, we use points
to represent the actions and arrows to illustrate the
order of actions. The points are placed in the regions
where the actions occurred. Therefore, an offensive
tactic is visualized as a sequence of actions on the
pitch (left part in Fig. 6 B).
Fig. 6. Visualization of tactics. (A) illustrates the simplification of offen-
Defensive tactic. A defensive tactic is a group of
sive tactics. (B) shows the design of the tactic diagram. (C) and (D) are defensive actions that occurred in the same spatial
icons for offensive and defensive tactic categories [8]. region. We use a point in the particular region to
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
5186 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 29, NO. 12, DECEMBER 2023

represent the spatial region of a defensive tactic

(right part in Fig. 6 B).

Icon. Our experts state that the offensive tactics could
be classified as corner, simple free-kick, and simple
pass, and the defensive tactics as tackle, interception,
and foul. It is difficult to encode multiple tactic cate-
gories with color or other visual channels. Therefore,
we use icons to represent different tactic categories
[8] (Figs. 6 C and 6 D). The tactic category is decided
by the type of the tactic’s first action. Thus, we put
the icon on the first action of a certain tactic to indi-
cate the category it belongs to.
Justification. Previous work has provided representations
for spatio-temporal events, such as sequence-based visuali-
zation and heatmap, but they cannot help users effectively
Fig. 7. The explanation component. (A) includes the switch buttons to
perceive the relative positions and the order of actions in switch among different modes. (B) illustrates the folded offensive inter-
the same view [59], [60], [61]. Thus, we decide to place the actions under the receiver mode. (C) Shows the offensive interactions
actions on their occurred positions in the pitch and link unfolded by forwards under the receiver mode.
them with arrows to show the action order. Moreover,
coaches concern about how to specify tactics to attack the switch button on the top right corner of the view.
defensive weakness and defend against the attacking of Meanwhile, the tactics of the opponent team will
the opponent. Such confrontation involves offensive and also be switched.
defensive tactics from different teams in the same spatial
Selecting with categories. Users can click the icons in
region. Thus, we apply the same pitch division to offensive the distribution bar charts to select categories of tac-
and defensive tactics to facilitate the matching of confron- tics that appearing in the list (Figs. 5 A1 and 5 A2).
tation tactics.
Sorting and filtering. Users can sort the tactics by one
Confrontation Tactic List. The confrontation tactic list of the usage and success rates, and filter by the other
includes two sub-lists that correspond with the offensive and rate.
defensive tactics of both teams (Figs. 5 A1, 5 A2, 5 A3, and 5
Selecting tactics. Users can click the tactics of the
A4). When users focus on the offensive tactics of the target opponent team in the left list to filter the confronta-
team, the defensive tactics of the opponent team are shown tion tactics of the target team in the same spatial
simultaneously, or vice versa. We place two sub-lists side by region. Then, the tactics of the target team in the
side to represent the confrontation between two teams. right list can be selected for different player roles
The usage and success rates are widely used in tactic style (Figs. 5 A5 and 5 A6).
analysis. The usage rate is calculated by dividing the number
Unfolding details. Users can hover on the icons and the
of all phases by the total number of phases in the tactic, and the bar charts to show the tactic category that the icon rep-
success rate is defined as dividing the number of phases in the resents and the value of usage rate or success rate.
tactic by the number of phases that created a shot. The usage
and success rates could indicate the tactical style and the
strength and weakness of a team, respectively. 5.3 Player View
The distributions of the usage rate and success rate After selecting tactics that are expected to be used, users can
among tactic types of the team and its opponent are most generate the optimal lineup under current tactical prefer-
important to the understanding of the tactical styles of both ence in the player view. The player view consists of a lineup
teams (T1). Therefore, each sub-list contains distribution edit board (Fig. 5 B1), a candidate player list (Fig. 5 B2), and
bar charts to present the usage rate (the gray bars) and suc- an explanation component (Fig. 7). In this view, users can
cess rate (the colored bars) of all tactic categories as an over- explore player performances and select players included or
view of tactic style (Figs. 5 A1 and 5 A2). The detailed excluded in the lineup in the candidate player list (P1).
tactics are listed under the distribution bar charts. Coaches Then, users can obtain the generated lineup and further
usually pay attention to the usage rates and success rates to adjust it in the lineup edit board. In the explanation compo-
choose confrontation tactics when facing a given opponent nent, users can explore the interactions among the players
(T2). Thus, each row of the list contains the tactic diagram, to explain why the players are selected by the generated
the usage rate (the gray bars), and the success rate (the col- lineup (P2).
ored bars) (Figs. 5 A5 and 5 A6). The usage rate and the suc- Candidate Player List. The candidate player list provides a
cess rate in the list are also encoded by bar charts for sortable list of player metrics for users to specify core play-
efficient ranking and filtering. ers who need to be included and players who cannot appear
Interaction Tools. Users can interact with the tactic view in in the match (P1) (Fig. 5 B2). We present the players with the
various ways. The follow are some commonly used tools same role in the list because coaches need to compare them
provided by the system. when selecting core players. Each item in the sortable list
consists of an icon that represents the player, and player

Switching tactics. Users can switch between offensive metrics including the individual performance value, the
and defensive tactics of the target team with the total offensive interaction value, and the total defensive
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
CAO ET AL.: TEAM-BUILDER: TOWARD MORE EFFECTIVE LINEUP SELECTION IN SOCCER 5187

interaction value (Fig. 5 B4). All the metrics are symbolized
Switching player roles. Users can click the switch but-
as bar charts for simplicity and effective comparison. The ton of player roles to switch the player role that
maximum length of the bar charts encodes the learned occurs in the candidate player list or the explanation
weight of the metric. The icons of the players selected by component.
users are highlighted with black borders, and those of the
Adjusting players. Users can drag unselected players
excluded players are indicated with the less-saturated color. in the candidate player list to the icons in the lineup
Players in the list can be sorted by each kind of metric editor board to replace players according to their
because coaches expect to select players according to differ- own opinions.
ent criteria.
Lineup Edit Board. The lineup edit board shows the lineup
result solved by the optimization equation or adjusted by 5.4 Lineup View
users (Fig. 5 B1). As the formation is an essential constraint With multiple generated solutions in the player view, users
when coaches select players, we divide the players in the can further identify the most suitable lineup in the lineup
lineup by their roles in formation, i.e., goalkeeper, defender, view with visual comparison. The lineup view includes a
midfielder, and forward, and place the players with the candidate lineup list (Fig. 5 C1) and lineup thumbnails
same role together (Fig. 5 B3). The lineup result is illustrated (Fig. 5 C2) for effective decision-making (L1, L2). In this
by actual positions in the formation for easy comprehension. view, users can sort lineups in the candidate lineup list to
Explanation Component. The explanation component shows decide on the most suitable lineup, and select lineups to
the detailed interaction values of each player (P2) (Fig. 7). obtain the details from lineup thumbnails for further
The interaction values among the players are represented comparison.
by a matrix and encoded by the area of rectangles. The Candidate Lineup List. The candidate lineup list records
players of the target team are illustrated by the rows of the the solutions and provides a flexible comparison according
matrix and are ranked by the total value of the row (Fig. 7 to different criteria to choose the best lineup (Fig. 5 C1).
B2). The matrix contains two modes to show offensive and Each row of the list contains four components: the predicted
defensive interaction values (Fig. 7 A1). The columns of match result, the confidence score, the predicted statistic
the matrix indicate the teammates in the offensive interac- indicators, and the formation icon [7] (Fig. 5 C4). We use bar
tion mode, or the opponents in the other mode. In the charts to present the lineup metrics to facilitate comparison.
offensive interaction mode, the rows and columns present The predicted match result is the probability of winning/
the passers and the receivers during the interactions. The drawing/losing the match. The confidence score is defined
players on the columns are initially summarized by player as the normalized average appearance time in the historical
roles and can be unfolded to show the players in a certain data of all players in the lineup, which could reflect the reli-
player role (Figs. 7 B1 and 7 C1). When the players on the ability of the lineup. The predicted statistic indicators
columns are unfolded, the individual performance and include offensive indicators (i.e., passing number and the
total defensive interaction values of each player are also success rate of passing), scoring indicators (i.e., shot number
illustrated by bars (Fig. 7 C3). Those players selected in the and goal number), and defensive indicators (i.e., intercep-
lineup are highlighted with the more saturated color for tion number and tackle number). Both the match result and
their icons (Figs. 7 B2 and 7 C1). the statistical indicators are predicted by the features of all
Justification. We present the relationships among the players in the lineup with a random forest approach.
players with a matrix, rather than a node-link diagram, to Lineup Thumbnails. We design lineup thumbnails to pres-
avoid visual clutter. Coaches require to compare the ent the details of candidate lineups (Fig. 5 C2). When users
selected players with the unselected players in the same select two lineups that need to be compared from the candi-
role to explain why the players are selected by the model. date lineup list, the thumbnails of the two lineups will be
Thus, rather than presenting all candidate players, the row presented on the right of the list. The player icons in the
and the column of the matrix only illustrate players in the lineup thumbnails are placed according to the player posi-
same role. tions in the team formation. We add an arc around each
Interaction Tools. Users can interact with the player view player icon to encode the confidence score of the player. For
in various ways. The follow are some commonly used tools convenient comparison among the lineups, those players
provided by the system. who are not picked by all selected lineups are highlighted

Specifying players. Users can click the player icons in with black borders in the lineup thumbnails.
the candidate player list to specify the players who Interaction Tools. Users can interact with the lineup view
need to be included and those who cannot appear in in various ways. The follow are some commonly used tools
the lineup. provided by the system.

Adjusting constraints. Users can drag the sliders in the
Sorting lineups. Users can sort lineups by each metric
lineup edit board to adjust the maximum and mini- for effective multi-criteria decision-making.
mum numbers of players of each role (Fig. 5 B3).
Switching indicators. Users can switch between differ-

Solution operations. Users can click the generate team ent groups of statistic indicators to evaluate whether
button to obtain the optimal lineup under the current the lineup meets their expectations (Fig. 5 C3).
input of tactics and players. In addition, they can delete
Unfolding details. Users can click lineups in the candi-
the current lineup with the clear button and add the date lineup list to show the details with lineup
current lineup to the lineup view with the add button. thumbnails.
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
5188 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 29, NO. 12, DECEMBER 2023

6 SYSTEM EVALUATION
In this section, we describe two case studies on the use of
our system in selecting and understanding lineups. After
the case studies, we also summarize the feedback of the
experts.

6.1 Case Studies

We invited two experts to conduct two case studies on the
use of our system. Our first expert is a senior analyst in
sports science (EA), and our second expert is a senior soccer
coach (EB). The cases used are from the soccer event dataset
in our analysis with the matches from the five top European
leagues in the 2017/18 season [56]. At the beginning of each
case study, we introduced the design and interaction tools
of our system. After the experts were familiar with our sys-
tem, they tried to select and analyze the lineup of a team of
interest. We recorded the analytical process and comments
of each expert. The findings and the activities of the two
case studies are summarized as follows.

6.1.1 Lineup Explanation: Why a Player is Selected by

the Automatically Generated Lineup?
This insight was gained from the first case study conducted
by EA. This case study is about exploring proper lineups for
Barcelona when facing Real Madrid. These two teams are Fig. 8. Analysis pipeline in Case 1. (A) presents the tactic specifications.
(B) shows the lineup generated for Barcelona when facing Real Madrid.
top teams in Spanish La Liga, and their competitions get sig- (C) is the comparison between the generated and the adjusted lineup.
nificant attention among soccer analysts. As mentioned by
EA, in the 2017/18 season, Barcelona won the championship
of La Liga and recorded one win (0:3) and one draw (2:2) in After the tactics were specified, EA moved to the player
matches against Real Madrid. Therefore, EA was interested view to generate and explore suitable lineups for Barcelona.
in generating lineups under such a scenario. EA first explored core players who needed to be included in
Before exploring the lineups, EA selected Barcelona as the lineup. In the candidate player list, EA noticed that the
the target team and Real Madrid as the opponent team, and learned weights of individual performance value and offen-
investigated the tactics of Barcelona when facing Real sive interaction value were larger than defensive interaction
Madrid in the tactic view. EA first focused on the defensive value, meaning that these two features contribute most to
tactics of Barcelona and the offensive tactics of Real Madrid. the match results (Fig. 5 B4). Thus, EA chose L. Messi as the
In the distribution bar charts, EA filtered the offensive tac- core player in the lineup because of his excellent individual
tics of Real Madrid with passing because those tactics were performance and offensive interactions with his teammates.
used most frequently and the success rate is relatively high Then, EA obtained the optimal lineup provided by the
(Fig. 5 A1). EA also filtered the defensive tactics of Barce- model and found that most players in the lineup met the
lona with tackle and interception for the same reason (Fig. 5 expectations, including powerful forwards such as L.
A2). After filtering tactic categories, EA switched to detailed Su
arez and outstanding midfielders like Iniesta (Fig. 8 B).
tactics of both teams. Specifically, EA sorted the tactics by Based on the lineup result, EA moved to the explanation
success rate and filtered out the tactics whose usage rates component to explain why the players were selected by the
are less than 5% to examine important tactics. As for Real model. EA mainly focused on the defenders because of the
Madrid, EA noticed that the passing tactics with high suc- tactical preference for defense. On the defensive aspect, EA
cess rates are mainly on the left-wing (Fig. 5 A3). Thus, EA noticed that Piqu
e, Sergi Roberto, S. Umtiti, and Jordi Alba
selected the first offensive tactic in the left list to filter those were ranked in the top four of all the defenders (Fig. 8 B1).
defensive tactics of Barcelona that are on the same spatial To be mentioned, Piqu
e, Sergi Roberto, and S. Umtiti were
regions (Fig. 8 A1). Thereafter, EA selected the defensive well-performed on defending the forwards and midfielders
tactic of Barcelona with the highest success rate and of the opponent team, and Jordi Alba was outstanding in
assigned it to defenders and midfielders based on the tacti- defending the opposing midfielders (Figs. 8 B3 and 8 B4).
cal preference for defense (Fig. 8 A2). Similarly, when speci- On the offensive aspect, EA found that Jordi Alba and Sergi
fying the offensive tactics of Barcelona, EA filtered the Roberto could pass threatening balls to their teammates,
offensive tactics of Barcelona by the least effective defensive especially forwards and midfielders (Fig. 7 B2). Meanwhile,
tactic of Real Madrid, the central-backfield defense (Fig. 8 Piqu
e and S. Umtiti were able to pass valuable balls to other
A3). After the filtering, EA assigned the most successful defenders to increase the goal chance. As for a detailed eval-
offensive tactic in the right list, mid-attack, to the forwards uation, EA unfolded all forwards and found that Jordi Alba
of Barcelona (Fig. 8 A4). and Sergi Roberto could create more opportunities for L.
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
CAO ET AL.: TEAM-BUILDER: TOWARD MORE EFFECTIVE LINEUP SELECTION IN SOCCER 5189

Messi and L. Su
arez, who were also selected in the lineup S. Umtiti rather than T. Vermaelen because of his more
and could rely on their excellent individual performances to effective interactions with excellent midfielders who could
score (Fig. 7 C2). EA summarized that the four defenders act as tactic organizers in matches, leading to more passes
were selected not only because they could prevent the and a higher goal chance.
attacking from the opponents, but also pass critical balls to To inspect the detailed confidence score of each player,
their teammates to score a goal. EA further clicked to display the thumbnails of the two line-
EA mentioned that the lineup recommended by the model ups. EA found that the confidence scores of all the players
had one different player compared with the lineup selected who were selected by both lineups are more than a half
by the coach of Barcelona in the match that defeated Real (Fig. 8 C3). This indicated that those players were per-
Madrid. Specifically, the model selected the defender S. formed stable and trusted to obtain extensive chances to
Umtiti, while the coach chose another defender T. Vermaelen. appear in matches. As for the two different players, EA
EA wondered why the lineup selection model chose S. Umtiti found that S. Umtiti attended more time in matches than T.
rather than T. Vermaelen. Thus, EA compared those two play- Vermaelen. After comparing the two lineups, EA concluded
ers in the explanation component. On the defensive aspect, that the model-generated lineup L1 would be more appro-
EA found that S. Umtiti performed more effective defense on priate for the match with Real Madrid than the coach-
all roles of the opponent players than T. Vermaelen, especially selected lineup L2.
on forwards (Fig. 8 B4). This suggested that S. Umtiti, rather Through this process, EA adjusted the generated lineup
than T. Vermaelen, could steal threatening balls from the from the perspective of coaches and discovered the differen-
opposing forwards. On the offensive aspect, EA noticed that ces between the expected performances of the two lineups.
S. Umtiti created more valuable passes to midfielders, while Such discoveries indicate that our system can provide more
T. Vermaelen interacted more closely with defenders (Fig. 8 effective lineups for real matches.
B5). This indicated that S. Umtiti was more likely to create
chances through midfielders, while T. Vermaelen tended to
pass to other defenders to start a build-up. EA further 6.1.3 Lineup Comparison: Which Lineup is the Best?
unfolded all midfielders and defenders for a detailed analysis.
This insight was gained from the second case study con-
EA found that S. Umtiti passed more valuable balls to the
ducted by EB. This case study is about choosing a lineup for
skilled midfielder Iniesta, and T. Vermaelen interacted effec-
Manchester United when competing with Manchester City.
tively with Sergi Roberto, the defender who was good at creat-
These two teams were the top two in the 2017/18 season of
ing chances for forwards. EA concluded that S. Umtiti was
the English Premier League. As mentioned by EB, Manchester
selected by the model due to better performance on defense
and more directly offensive interactions with midfielders. United got one lose (1:2) and one win (3:2) when facing
Manchester City in the season. EB was interested in improv-
Through this process, EA generated a lineup for Barcelona
ing the lineup of Manchester United with our system.
under a combination of tactics and explained why those play-
EB began with the exploration from strategy specification
ers in the lineup were recommended by the system.
in the tactic view. First, EB filtered the passing tactics of
Manchester City and the interception tactics of Manchester
6.1.2 Adjusting Players: Which Player is More United, selected the most effective passing tactic of Man-
Suitable? chester City, and assigned the most successful interception
tactic on the same spatial region to the defenders (Figs. 9 A1
This insight was also gained from the first case study. After
and 9 A3). Afterwards, EB focused on the interception tac-
the process of the first insight, EA required to know whether
tics of Manchester City and the passing tactics of Manches-
the lineup generated in the system could be better than the
lineup in the real match. Therefore, EA replaced S. Umtiti ter United, selected the least used interception tactic of
Manchester City, and assigned the corresponding passing
with T. Vermaelen in the generated lineup and added the
tactic to the midfielders and forwards due to the tactical
two lineups to the lineup view for further comparison.
preference for offense (Figs. 9 A2 and 9 A4). Then, EB
We refer to the lineup including S. Umtiti as L1 and the
turned to the player view to get lineups for Manchester
lineup including T. Vermaelen as L2. In the candidate
United. EB chose P. Pogba as the core player due to his out-
lineup list, EA first examined the two lineups on the aspect
standing offensive interaction with his teammates (Fig. 9 B),
of the confidence score and the predicted match result to
and obtained three lineups under the current tactic selec-
identify the better lineup. EA found that the confidence
tion. The first lineup was obtained directly without addi-
score of L1 was higher than that of L2, meaning that the pre-
tional operations. The second lineup was constrained by a
dicted performance of L1 is more robust than that of L2
4-3-3 formation, which was often used by Manchester
(Fig. 8 C2). Meanwhile, both the predicted winning rate and
the chance of not losing the match of L1 were higher than United through the season. The third lineup was adjusted
by EB based on the second lineup. More specifically, EB
those of L2 (Fig. 8 C1). Based on the two aspects, EA sum-
replaced an offensive midfielder, J. Lingard (Fig. 9 B2), with
marized that L1 would be a better choice for the real match.
H. Mkhitaryan (Fig. 9 B1), another offensive midfielder with
Then, EA focused on the predicted statistical indicators to
better offensive interactions with the teammates, to find out
discover the details of the two lineups. EA switched among
whether a lineup with cooperative but less skilled players
the three kinds of indicators and noticed that L1 would per-
for Manchester United would perform better. EB also gener-
form better on passing number and goal number, while
ated another lineup with a different tactic selection by
other indicators were not different obviously. This con-
replacing the tactic for the midfielders with the most
firmed the previously gained insight that the model selected
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
5190 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 29, NO. 12, DECEMBER 2023

(Figs. 9 E1 and 9 E2). As for the statistical indicators, EB

found that both the two scoring indicators of L1 were higher
than L2 (Fig. 9 E5). This confirmed the hypothesis of EB that
L1 was a typical offensive formation 3-4-3, and the offensive
midfielder M. Fellaini was well-performed in attacking than
the defender P. Jones. EB further switched to the other two
groups of indicators and found that L1 would create slightly
fewer passes (Fig. 9 D) and more tackles than L2 (Fig. 9 E6).
It also corresponded to the knowledge of EB that L1 would
be more suitable for the high-press strategy, while L2 could
be adapted for the build-up strategy. EB concluded that if
Manchester United took offensive strategies, L1 could be
the most suitable lineup, and L2 would be the most proper
lineup if more stable strategies were employed.
Through this process, EB obtained four different lineups
for Manchester United with different tactic combinations
when facing Manchester City. EB also gained insights about
comparing those lineups and choosing the most suitable
one under different scenarios.

6.2 Expert Interview

We interviewed the two experts who conducted the case
studies (EA and EB) and the other two new experts (EC and
ED) to evaluate the usefulness and effectiveness of our sys-
tem. Both the two new experts are doctoral students in
physical education and are experienced in soccer data anal-
ysis. We interviewed those experts respectively and col-
lected their feedback during the interview.
Fig. 9. Analysis pipeline of Case 2. (A) presents the tactic specifications.
(B) and (C) illustrate the lineup adjustments. (D) Shows the four lineups Procedure. The interview procedure for the two new
generated for Manchester United. (E) Presents the multi-criteria compar- experts (EC and ED) contains four steps. First, we briefly
ison of lineups. introduced the metrics and the model used in our soccer
lineup selection system. Then, we presented the visual
successful free-kick tactic as it was another essential attack- design and interactions to the experts through a usage
ing method for Manchester United (Fig. 9 C). scenario. We also let the experts freely explore the sys-
EB added those four lineups (referred as L1, L2, L3, and tem and answered their questions to help them become
L4) into the lineup view to examine the overall performances familiar with the system. Afterward, the experts were
with multi-criteria comparison among different indicators required to select the most appropriate lineup for Barce-
(Fig. 9 D). EB first clicked the table headers of the winning lona when facing Real Madrid. We recorded their analy-
rate and the appearance time to identify the most appropri- sis procedures during this step. Finally, we asked the
ate lineup. EB found that L1 and L2 ranked the top two under experts whether each design requirement (i.e., T1-T2, P1-
the winning rate (Figs. 9 D1 and 9 D2), and all the three line- P2, L1-L2) has been solved by the visual design, whether
ups under the first tactic selection were higher than L4 under there are potential issues in the current design, and
the appearance time (Fig. 9 D3). Such comparison results whether the system is sufficient to support a more effec-
suggested that the first tactic selection would be more appro- tive and efficient lineup selection process. After EA and
priate for matches against Manchester City. Besides, EB men- EB finished the case study, we also asked those ques-
tioned that L2 was better than L3 on both the winning rate tions to gain their feedback.
and appearance time (Fig. 9 D3), indicating that the lineup System Usability. Generally, all the experts were satis-
with players only outstanding in passing threatening balls fied with our system and felt that it could help soccer
would perform worse than balanced lineups under the first coaches select the lineup effectively. Compared with
tactic specification. Furthermore, EB noticed that L1 was the directly selecting players by experience, the coaches could
best in winning the match but was ranked third in robustness specify their tactical preferences based on the tactic detec-
(Fig. 9 D1). Meanwhile, the expected performance of L2 was tion results and generate lineups interactively with our
the most robust and ranked second in the winning rate system. EA commented that “Coaches usually select lineups
(Fig. 9 D2). It meant that coaches are required to make a by their own opinions on player performance, which is labori-
trade-off between the two lineups. Thus, EB turned on the ous and subjective. Such a system could provide them an effi-
thumbnails of L1 and L2 for a detailed comparison. cient method to select lineups based on historical data and
Through the lineup thumbnails, EB found that L1 evaluate them by expected performance”. The experts also
selected the offensive midfielder M. Fellaini, who per- thought highly of our lineup selection model because it
formed well but seldom appeared in the historical matches, considers the relationships between players and could
and L2 chose the more experienced defender P. Jones provide more reasonable lineup results. EB indicated that
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
CAO ET AL.: TEAM-BUILDER: TOWARD MORE EFFECTIVE LINEUP SELECTION IN SOCCER 5191

“Most of the players selected by the system matched my expect- Lessons Learned. We have learned two lessons through the
ations. Besides, the system could recommend new lineups I research. The first lesson is from the evaluation of defensive
have not thought of previously, which inspired me to explore interactions. Defending in soccer matches is difficult to eval-
whether a known lineup could be improved through lineup uate because it cannot directly create goals. During the
adjustments and the prediction of lineup performance”. They meetings, the experts indicated that the defensive interac-
also mentioned that with the system, coaches could tions mainly contribute to preventing the goal of the oppo-
understand why the players were selected by the model nent team and increasing the goal chance of the team. In
and decide whether apply the lineup in real matches. this way, the defensive interaction between two players
Visual Design. All the experts agreed that each design from different teams can be evaluated through the change
requirement has been fulfilled by our visual design. Specifi- of goal probability of the two teams. The second lesson is
cally, ED appreciated the visual design for tactic specifica- from the visualization of offensive tactics. Compared with
tion, “... the confrontation tactic list is helpful in discovering placing all phases in the tactic on the soccer pitch (the left
essential tactics, and the presentation of tactics are effective for part of Fig. 6 A), an aggregated sequence of the tactic (the
identifying the tactic at a glance”. EB and EC liked the candi- right part of Fig. 6 A) could help users clearly identify how
date player list and mentioned that it could be particularly the ball was passed through the pitch. We further divide the
useful in the scenario of selecting the lineup for a real pitch into nine regions to match offensive tactics with defen-
match. “The comparison and ranking would be effective when sive tactics (Fig. 6 B). The case study has shown that the
selecting the core players” (EC). “Injuries and match bans com- aggregated sequence and pitch division are effective in the
monly occur in lineup selection. I did not use the function of exploration of confrontation tactics. Such designs could also
removing players in the case study because of the lack of data. I be generalized to other team sports such as basketball and
can remove those players when selecting lineups in real matches” ice hockey.
(EB). As for the explanation component, EA was impressed Limitations. There are some limitations in this research.
by the matrix-based design, “It is convenient to find player First, our model does not include special factors that may
pairs with successful interactions by the matrix”. ED also influence a lineup, such as the physical and mental condi-
favored the multi-criteria comparison of the lineups, “... it tions of players, due to lack of relevant data. These factors
can help me make a trade-off among multiple lineups and quickly should be considered when coaches need to select more reli-
find the one that I desired”. able lineups for the match. We plan to collect such data and
Suggestions. The experts had some suggestions for integrate them into the measurement of player teamwork to
improvement. EA focused on the lineup selection model improve the reliability of our model. Second, the player role
and hoped to add movement data of players without the constraints in our model are limited to the four basic roles
ball for a more comprehensive evaluation of player interac- (i.e., goalkeeper, defender, midfielder, forward). The pre-
tions. EB and EC considered the interactions in the player ferred side indicated by the detailed role (e.g., left-winger,
candidate list and believed that a search bar could be more center forward) is not considered. These fine-grained con-
effective for finding players who cannot appear in the straints should be considered if more detailed specification
match. ED concentrated on the system design and sug- of lineups is required. We will develop more sophisticated
gested providing a record list for specifications and con- models to improve the player role constraints to deal with
straints for all lineups. the preferred side as the future work.

7 DISCUSSION 8 CONCLUSION
Significance. Lineup selection plays an important role in In this article, we present a method for visual analytics of
team sports, such as basketball and soccer. Existing work soccer lineup selection. Collaborating with the domain
on automatic soccer lineup selection is weak in incorpo- experts, we characterize the problem and propose a new
rating the inputs from coaches and supporting interac- teamwork-based model to integrate the factors preferred by
tive exploration of lineup possibilities. In this research, coaches (e.g., spatial regions, opponent information) into
we develop an approach to address these challenges lineup selection. Based on the model, we develop a web-
through modeling player performances and developing based visual analytics system, Team-Builder, to allow
an interactive visual analytics system for lineup genera- coaches to interactively generate, explain, and compare line-
tion, explanation, and comparison. Our approach allows ups produced by an automatic model, and integrate their
coaches to steer the selection and choose the most appro- own opinions into lineup selection.
priate lineup based on their own opinions and the oppo- In the future, our work can be improved in several ways.
nent information. One direction is to enhance our model by considering more
Our approach can be generalized to other team sports with data types (e.g., movement of players without the ball, inter-
the requirements on personal and team skills similar to soccer, actions among multiple players in a period of time) so that
such as basketball, ice hockey, and baseball. Our model is coaches can evaluate the performances of players and their
intrinsically composed of the summarization of individual interactions more comprehensively. With these data, our
performance of players and player interactions in matches, lineup metrics can also be integrated into other analysis
and can be converted for other team sports by incorporating tasks such as evaluating the change of team formation.
appropriate evaluation metrics of individual performance Another direction is to extend our work to other team sports
and player interactions. Our visual design can also be easily such as basketball, ice hockey, and baseball by adapting our
modified based on the characteristics of sports. models, algorithms, and visualization tools based on the
Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
5192 IEEE TRANSACTIONS ON VISUALIZATION AND COMPUTER GRAPHICS, VOL. 29, NO. 12, DECEMBER 2023

characteristics of individual sports. Besides, we also plan to [22] W. Wang, J. Zhang, X. Yuan, and S. Liu, “MatchOrchestra: A gen-
eralized visual analytics for competitive team sports,” J. Vis.,
improve the system usability based on the long-term expert vol. 19, no. 3, pp. 515–528, 2016.
feedback for the real lineup selection scenario. [23] J. P. Ono, C. A. Dietrich, and C. T. Silva, “Baseball timeline: Sum-
marizing baseball plays into a static visualization,” Comput. Graph.
Forum, vol. 37, no. 3, pp. 491–501, 2018.
ACKNOWLEDGMENTS [24] M. Lage, J. P. Ono, D. Cervone, J. Chiang, C. A. Dietrich, and C. T.
We would like to thank Jiahui Chen and Rui Yang from the Silva, “StatCast dashboard: Exploration of spatiotemporal base-
ball data,” IEEE Comput. Graph. Appl., vol. 36, no. 5, pp. 28–37,
State Key Lab of CAD&CG, Zhejiang University, for their Sep./Oct. 2016.
contribution to system development. [25] C. A. Dietrich, D. Koop, H. T. Vo, and C. T. Silva, “Baseball4D: A
tool for baseball game reconstruction and visualization,” in Proc.
IEEE Conf. Vis. Analytics Sci. Technol., 2014, pp. 23–32.
REFERENCES [26] B. Moon and R. Brath, “Bloomberg sports visualization for pitch
[1] B. H. Boon and G. Sierksma, “Team formation: Matching quality analysis,” in Proc. Workshop Sports Data Vis., 2013, pp. 11–19.
supply and quality demand,” Eur. J. Oper. Res., vol. 148, no. 2, [27] T. Polk, D. J€ackle, J. H€außler, and J. Yang, “CourtTime: Generating
pp. 277–292, 2003. actionable insights into tennis matches using visual analytics,” IEEE
[2] H. Ruiz, P. Power, X. Wei, and P. Lucey, ““The Leicester City fairy- Trans. Vis. Comput. Graphics, vol. 26, no. 1, pp. 397–406, Jan. 2020.
tale?”: Utilizing new soccer analytics tools to compare performance [28] S. Ye et al., “ShuttleSpace: Exploring and analyzing movement
in the 15/16 & 16/17 EPL seasons,” in Proc. 23rd ACM SIGKDD Int. trajectory in immersive visualization,” IEEE Trans. Vis. Comput.
Conf. Knowl. Discov. Data Mining, 2017, pp. 1991–2000. Graphics, vol. 27, no. 2, pp. 860–869, Feb. 2021.
[3] UEFA, “EURO 2020 technical report,” 2021. [Online]. Available: [29] J. Wu, D. Liu, Z. Guo, Q. Xu, and Y. Wu, “TacticFlow: Visual ana-
https://uefatechnicalreports.com/uefa-euro-2020 lytics of ever-changing tactics in racket sports,” IEEE Trans. Vis.
[4] E. Ozceylan, “A mathematical model using AHP priorities for soc- Comput. Graphics, vol. 28, no. 1, pp. 835–845, Jan. 2022.
cer player selection: A case study,” South Afr. J. Ind. Eng., vol. 27, [30] J. Lan, J. Wang, X. Shu, Z. Zhou, H. Zhang, and Y. Wu,
no. 2, pp. 190–205, 2016. “RallyComparator: Visual comparison of the multivariate and
[5] S. A. F. Salles, H. R. M. da Hora, M. E. J
unior, A. C. de Souza spatial stroke sequence in table tennis rally,” J. Vis., vol. 25, no. 1,
Gomes dos Santos, and A. Shimoya, “Operations research contri- pp. 143–158, 2022.
butions for football teams formation: A systematic review,” Pes- [31] P. A. Legg et al., “MatchPad: Interactive glyph-based visualization
quisa Operacional, vol. 39, no. 2, pp. 277–293, 2019. for real-time sports performance analysis,” Comput. Graph. Forum,
[6] G. L. Andrienko et al., “Constructing spaces and times for tactical vol. 31, no. 3, pp. 1255–1264, 2012.
analysis in football,” IEEE Trans. Vis. Comput. Graphics, vol. 27, [32] H. Pileggi, C. D. Stolper, J. M. Boyle, and J. T. Stasko, “SnapShot:
no. 4, pp. 2280–2297, Apr. 2021. Visualization to propel Ice hockey analytics,” IEEE Trans. Vis.
[7] Y. Wu et al., “For Vizor: Visualizing spatio-temporal team forma- Comput. Graphics, vol. 18, no. 12, pp. 2819–2828, Dec. 2012.
tions in soccer,” IEEE Trans. Vis. Comput. Graphics, vol. 25, no. 1, [33] Y. Ishikawa and I. Fujishiro, “TideGrapher: Visual analytics of tac-
pp. 65–75, Jan. 2019. tical situations for rugby matches,” Vis. Inform., vol. 2, no. 1,
[8] X. Xie et al., “PassVizor: Toward better understanding of the pp. 60–70, 2018.
dynamics of soccer passes,” IEEE Trans. Vis. Comput. Graphics, [34] C. Perin, R. Vuillemot, and J.-D. Fekete, “A table!: Improving tem-
vol. 27, no. 2, pp. 1322–1331, Feb. 2021. poral navigation in soccer ranking tables,” in Proc. CHI Conf. Hum.
[9] A. Cao et al., “MIG-Viewer: Visual analytics of soccer player Factors Comput. Syst., 2014, pp. 887–896.
migration,” Vis. Inform., vol. 5, no. 3, pp. 102–113, 2021. [35] C. Perin, J. Boy, and F. Vernier, “Using gap charts to visualize the
[10] D. McEwan and M. R. Beauchamp, “Teamwork in sport: A theo- temporal evolution of ranks and scores,” IEEE Comput. Graph.
retical and integrative review,” Int. Rev. Sport Exercise Psychol., Appl., vol. 36, no. 5, pp. 38–49, Sep./Oct. 2016.
vol. 7, no. 1, pp. 229–250, 2014. [36] A. Rusu, D. Stoica, and E. Burns, “Analyzing soccer goalkeeper
[11] D. W. Eccles and G. Tenenbaum, “Why an expert team is more performance using a metaphor-based visualization,” in Proc. 15th
than a team of experts: A social-cognitive conceptualization of Int. Conf. Inf. Vis., 2011, pp. 194–199.
team coordination and communication in sport,” J. Sport Exercise [37] A. Rusu, D. Stoica, E. Burns, B. Hample, K. McGarry, and R. Rus-
Psychol., vol. 26, no. 4, pp. 542–560, 2004. sell, “Dynamic visualizations for soccer statistical analysis,” in
[12] D. W. Eccles, “The coordination of labour in sports teams,” Int. Proc. 14th Int. Conf. Inf. Vis., 2010, pp. 207–212.
Rev. Sport Exercise Psychol., vol. 3, no. 2, pp. 154–170, 2010. [38] C. Perin, R. Vuillemot, and J.-D. Fekete, “SoccerStories: A kick-off
[13] J. Kuehn, “Accounting for complementary skill sets when evaluat- for visual soccer analysis,” IEEE Trans. Vis. Comput. Graphics,
ing NBA players’ values to a specific team,” in Proc. MIT Sloan vol. 19, no. 12, pp. 2506–2515, Dec. 2013.
Sports Analytics Conf., 2016, pp. 1–11. [39] J. L. S. Malqui, N. M. L. Romero, R. Garcia, H. Alemdar, and J. L.
[14] A. Z. Maymin, P. Z. Maymin, and E. Shen, “NBA chemistry: Posi- Comba, “How do soccer teams coordinate consecutive passes? A
tive and negative synergies in basketball,” Int. J. Comput. Sci. visual analytics system for analysing the complexity of passing
Sport, vol. 12, no. 2, pp. 4–23, 2013. sequences using soccer flow motifs,” Comput. Graph., vol. 84,
[15] R. Beal, N. Changder, T. J. Norman, and S. D. Ramchurn, pp. 122–133, 2019.
“Learning the value of teamwork to form efficient teams,” in Proc. [40] M. Stein et al., “Director’s cut: Analysis and annotation of soccer
AAAI Conf. Artif. Intell., 2020, pp. 7063–7070. matches,” IEEE Comput. Graph. Appl., vol. 36, no. 5, pp. 50–60,
[16] L. Bransen and J. van Haaren, “Player chemistry: Striving for a Sep./Oct. 2016.
perfectly balanced soccer team,” 2020, arXiv:2003.01712. [41] M. Stein et al., “Bring it to the pitch: Combining video and move-
[17] J. Fernandez-Navarro, L. Fradua, A. Zubillaga, P. R. Ford, and ment data to enhance team sport analysis,” IEEE Trans. Vis. Com-
A. P. McRobert, “Attacking and defensive styles of play in soccer: put. Graphics, vol. 24, no. 1, pp. 13–22, Jan. 2018.
Analysis of spanish and English elite teams,” J. Sports Sci., vol. 34, [42] L. Shao, D. Sacha, B. Neldner, M. Stein, and T. Schreck, “Visual-
no. 24, pp. 2195–2204, 2016. interactive search for soccer trajectories to identify interesting
[18] C. Perin, R. Vuillemot, C. D. Stolper, J. T. Stasko, J. Wood, and S. game situations,” Electron. Imag., no. 1, pp. 1–10, 2016.
Carpendale, “State of the art of sports data visualization,” Comput. [43] D. Sacha et al., “Dynamic visual abstraction of soccer movement,”
Graph. Forum, vol. 37, no. 3, pp. 663–686, 2018. Comput. Graph. Forum, vol. 36, no. 3, pp. 305–315, 2017.
[19] M. Du and X. Yuan, “A survey of competitive sports data visualiza- [44] G. L. Andrienko et al., “Visual analysis of pressure in football,”
tion and visual analysis,” J. Vis., vol. 24, no. 1, pp. 47–67, 2021. Data Mining Knowl. Discov., vol. 31, no. 6, pp. 1793–1839, 2017.
[20] A. Franks, A. Miller, L. Bornn, and K. Goldsberry, “Counterpoints: [45] H. Janetzko, D. Sacha, M. Stein, T. Schreck, D. A. Keim, and O.
Advanced defensive metrics for NBA basketball,” in MIT Sloan Deussen, “Feature-driven visual analytics of soccer data,” in Proc.
Sports Analytics Conf., 2015, pp. 1–8. IEEE Conf. Vis. Analytics Sci. Technol., 2014, pp. 13–22.
[21] A. G. Losada, R. Ther
on, and A. Benito, “BKViz: A basketball [46] V. Machado, R. Leite, F. Moura, S. Cunha, F. Sadlo, and J. L.
visual analysis tool,” IEEE Comput. Graph. Appl., vol. 36, no. 6, Comba, “Visual soccer match analysis using spatiotemporal posi-
pp. 58–68, Nov./Dec. 2016. tions of players,” Comput. Graph., vol. 68, pp. 84–95, 2017.

Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.
CAO ET AL.: TEAM-BUILDER: TOWARD MORE EFFECTIVE LINEUP SELECTION IN SOCCER 5193

[47] M. Stein et al., “Where to go: Computational and visual what-if Xiao Xie received the PhD degree in computer
analyses in soccer,” J. Sports Sci., vol. 37, no. 24, pp. 2774–2782, science from Zhejiang University. He is an assis-
2019. tant professor with the Department of Sports Sci-
[48] A. B. Santos, R. Theron, A. Losada, J. E. Sampaio, and C. Lago-Pe ence, Zhejiang University. His research interests
~ as, “Data-driven visual performance analysis in soccer: An
n include data visualization, visual analytics, and
exploratory prototype,” Front. Psychol., vol. 9, 2018, Art. no. 2416. human-computer interaction, with a focus on cre-
[49] L. Shaw and M. Glickman, “Dynamic analysis of team strategy in pro- ating novel visual analysis techniques for sup-
fessional football,” Barça Sports Analytics Summit, vol. 13, pp. 1–13, porting sports analysis.
2019.
[50] A. Bialkowski, P. Lucey, P. Carr, Y. Yue, S. Sridharan, and I. Mat-
thews, “Large-scale analysis of soccer matches using spatiotempo-
ral tracking data,” in Proc. IEEE Int. Conf. Data Mining, 2014,
pp. 725–730.
[51] J. Perl and D. Memmert, “Soccer analyses by means of artificial Hongyu Chen is currently a senior majoring in
neural networks, automatic pass recognition and Voronoi-cells: computer science and technology with Zhejiang
An approach of measuring tactical success,” in Proc. 10th Int. University. Her research interests include sports
Symp. Comput. Sci. Sports, 2016, pp. 77–84. data visualization and visual analytics.
[52] M. Tavana, F. Azizi, F. Azizi, and M. Behzadian, “A fuzzy infer-
ence system with application to player selection and team forma-
tion in multi-player sports,” Sport Manage. Rev., vol. 16, no. 1,
pp. 97–110, 2013.
[53] J. M. Merig
o and A. M. Gil-Lafuente, “Decision-making in sport
management based on the OWA operator,” Expert Syst. Appl.,
vol. 38, no. 8, pp. 10 408–10 413, 2011.
[54] R. Al-Shboul, T. Syed, J. Memon, and F. Khan, “Automated player
Xiaolong Zhang received the PhD degree from
selection for sports team using competitive neural networks,” Int.
J. Adv. Comput. Sci. Appl., vol. 8, no. 8, pp. 457–460, 2017. the School of Information, University of Michigan.
[55] T. Decroos, L. Bransen, J. V. Haaren, and J. Davis, “Actions speak He is an associate professor with the College of
louder than goals: Valuing player actions in soccer,” in Proc. 25th Information Sciences and Technology, Pennsyl-
ACM SIGKDD Int. Conf. Knowl. Discov. Data Mining, 2019, vania State University. His research interests
pp. 1851–1861. include using advanced interactive technologies
like multi-scale, virtual environments, and collab-
[56] L. Pappalardo et al., “A public data set of spatio-temporal match
events in soccer competitions,” Sci. Data, vol. 6, no. 1, pp. 1–15, oration to improve the understanding of large
2019. information structures and the design of interac-
[57] T. Decroos, J. V. Haaren, and J. Davis, “Automatic discovery of tive tools to support visually-guided navigation in
tactics in spatio-temporal soccer match data,” in Proc. 24th ACM large information space. Currently, he is working
SIGKDD Int. Conf. Knowl. Discov. Data Mining, 2018, pp. 223–232. on issues concerning multi-facet visualization of information structures
and multi-scale navigation.
[58] P. Power, H. Ruiz, X. Wei, and P. Lucey, ““Not all passes are cre-
ated equal”: Objectively measuring the risk and reward of passes
in soccer from tracking data,” in Proc. 23rd ACM SIGKDD Int.
Conf. Knowl. Discov. Data Mining, 2017, pp. 1605–1613. Hui Zhang received the bachelor’s degree from
[59] S. Latif and F. Beck, “Interactive map reports summarizing bivari- Zhejiang Normal University, the master’s degree
ate geographic data,” Vis. Inform., vol. 3, no. 1, pp. 27–37, 2019. from Beijing Sport University, and the PhD degree
[60] W. Zeng, C.-W. Fu, S. M. Arisona, S. Schubiger, R. Burkhard, and from the University of Potsdam, Germany. He has
K.-L. Ma, “A visual analytics design for studying rhythm patterns rich training and competition experiences and was
from human daily movement data,” Vis. Inform., vol. 1, no. 2, the level-1 table tennis player, international referee,
pp. 81–91, 2017. level-1 coach and the coach with A-level certificate
[61] B. Ni, Q. Shen, J. Xu, and H. Qu, “Spatio-temporal flow maps for in Germany. From 2003 to 2013, he was a profes-
visualizing movement and contact patterns,” Vis. Inform., vol. 1, sor of Shanghai University of Sport, the executive
no. 1, pp. 57–64, 2017. deputy dean of China Table Tennis College and
director of Key Lab of Technique and Tactic Diagno-
Anqi Cao received the bachelor’s degree in com- sis and Analysis, State General Administration of Sports. Since May, 2014,
puter science from Nankai University, in 2019. he works with the Department of Sports Science, College of Education,
Zhejiang University as a professor.
She is currently working toward the PhD degree
in computer science with the State Key Lab of
CAD&CG, Zhejiang University. Her research inter-
ests include sports data visualization and visual Yingcai Wu received the PhD degree in com-
analytics. puter science from the Hong Kong University of
Science and Technology. He is a professor with
the State Key Lab of CAD&CG, Zhejiang Univer-
sity. His main research interests include visual
analytics and information visualization, with
focuses on sports analytics and big data intelli-
Ji Lan is currently working toward the joint PhD gence. Prior to his current position, he was a
degree with the Yingcai Wu @State Key Lab of postdoctoral researcher with the University of
CAD&CG College of Computer Science, Zhejiang California, Davis from 2010 to 2012, and a
University, Hangzhou, China, and Jonathan ZHU researcher with Microsoft Research Asia from
@Department of Media and Communication, 2012 to 2015. For more information, please visit
CityU University, Hong Kong. His main research http://www.ycwu.org.
interests include visual analytics and computa-
tional social science. For more information,
please visit https://lanjizju.github.io/ " For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/csdl.

Authorized licensed use limited to: ShanghaiTech University. Downloaded on December 04,2024 at 05:31:22 UTC from IEEE Xplore. Restrictions apply.