identifying team style in soccer using formations...
TRANSCRIPT
Identifying Team Style in Soccer using FormationsLearned from Spatiotemporal Tracking Data
Alina Bialkowski1,2, Patrick Lucey1, Peter Carr1, Yisong Yue1,3, Sridha Sridharan2 and Iain Matthews11Disney Research, Pittsburgh, USA, 2Queensland University of Technology, Australia, 3California Institute of Technology, USA
Email: [email protected], {patrick.lucey, peter.carr, iainm}@[email protected], [email protected]
Abstract—To the trained-eye, experts can often identify a teambased on their unique style of play due to their movement,passing and interactions. In this paper, we present a methodwhich can accurately determine the identity of a team fromspatiotemporal player tracking data. We do this by utilizing aformation descriptor which is found by minimizing the entropyof role-specific occupancy maps. We show how our approach issignificantly better at identifying different teams compared tostandard measures (i.e., shots, passes etc.). We demonstrate theutility of our approach using an entire season of Prozone playertracking data from a top-tier professional soccer league.
I. INTRODUCTION
The question we ask in this paper is: given all the playerand ball tracking data of a team in a season, what team-based features can adequately discriminate a team’s behavior?In practice a human expert is able to do this, but it is verylabor intensive and is inherently subjective. Having a methodwhich can quantify these behaviors should be possible with theprevalence of spatiotemporal tracking data of player and ballmovement being captured in most professional sports (e.g., [1],[2]). However, this task is challenging due to the complexitiesin dealing with adversarial multi-agent trajectory data. A majorissue centers on the alignment of individual player trajectorieswithin a team setting which is a source of noise. In this paper,we align the data based on a role-based method which islearnt directly from data [3] to provide a formation descriptor.We show that using this approach, semantically meaningfulteam-based strategic features can be obtained which are highlypredictive of their identity. We compare this descriptor to otherfeatures including match statistics (e.g., shots, passes, fouls)and ball movement, and show that the formation descriptoris far superior in discriminating unique team characteris-tics (Fig. 1).
A. Related Work
With the recent deployment of player tracking systemsin professional sports, a recent influx of research has beenconducted on how to use such data sources. Most of thework has centered on individual player analysis. In basketball,Goldsberry [4] used player tracking data to rank the best shoot-ers in the NBA according to their shot location. Maheswaranet al. [5], [6] used the tracking data to analyze the best methodto obtain a rebound. Similarly, Wiens et al. [7] looked at howteams should crash the backboard to get rebounds. Recently,Lucey et al. [8] used tracking data to discover how teams
A B C D E F G H I J K L M N O P Q R S T
Team ID {StatisticsShots (on goal) 12(4)
Fouls 11
Corner kicks 8
Offsides 4
Time of possession 62%
Yellow cards 1
Red cards 0
Saves 3
FormationGame716, T1, GT Label = 4−1−4−1
12
34
56
7
8
9
10
Ball occupancyWEHAM
Average overall
Fig. 1. In this paper, based solely on match statistics, location of ballpossession (i.e., ball occupancy), and a formation descriptor, we can predict theidentity of soccer teams with high accuracy. We show the formation descriptoris the best discriminator of team style.
achieved open three-point shots. Bocskocksy et al. [9] re-investigated the hot-hand theory. Miller et al. [10] analyzedthe shot selection process of players using non-negative matrixfactorization. Cervone et al. [11] used basketball trackingdata to predict points and decisions made during a play.Carr et al. [12] used real-time player detection data to predictthe future location of play and point a robotic camera inthat location for automatic sport broadcasting purposes. Intennis, Wei et al. [13], [14] used Hawk-Eye data to predictthe type and location of the next shot. Ganeshapillai andGuttag [15] used SVMs to predict pitching in baseball whileSinha et al. [16] used Twitter feeds to predict NFL outcomes.
In terms of analyzing a team’s style of play, most workhas centered on soccer. Lucey et al. [17] used entropy mapsto characterize a team’s ball movement patterns using datafrom Opta [18]. This was followed by [19], which showedthat a team’s home and away style varied, highlighting thathome teams had more possession in the forward third aswell as shots and goals. Bialkowski et al. [20] examined therigidity of a team’s formation across a season and showedthat home teams tended to player higher up the pitch bothin offense and defense. Outside of the sporting realm, therehas been plenty of work focusing on identifying style. In theseminal work on separating style from content, Tenenbaumand Freeman [21] used a bilinear model to decouple theraw content for improved recognition on a host of differenttasks. More recently, Doersch et al. [22] used discriminativeclustering to discover the attributes that distinguished imagesof one city from another. They followed this work by exploringthe visual style of objects (e.g., cars and houses) and how theyvary over time [23]. The contribution of this paper is using aformation descriptor to identity the unique style of a team.
(a) (b) (c)
Fig. 2. (a) Given the player trajectory of each player during an entire half, we see that players continually swap positions. (b) Shown are the covariances ofplayer positions which again highlights the overlap. (c) Using our iterative approach (which is very similar to k-means with the constraint that at every frameeach detection requires a unique role), a role label is assigned to each player at the frame-level, allowing us to see the underlying structure of the team.
Statistic Frequency
Teams 20Games 375
Data Points 3.89MBall Events 721K
TABLE I. INVENTORY OF DATASET USED FOR THIS WORK.
II. DATA: PLAYER TRACKING IN SOCCER
For this work, we utilized an entire season of playertracking data from Prozone. The data consists of 20 teamswho played home and away, totaling 38 games for each teamor 380 games overall. Five of these games were omitted due toerroneous data files. We refer to the 20 teams using arbitrarylabels {A, B, . . . , T}. Each game consists of two halves,with each half containing the (x, y) position of every playerat 10 frames-per-second. This results in over 1 million data-points per game, in addition to the 43 possible annotated ballevents (e.g., passes, shots, crosses, tackles etc.). Each of theseball events contained the time-stamp as well as location andplayers involved. An inventory of the data is given in Table I.
III. DISCOVERING FORMATIONS FROM DATA
In sports, there exists a well established vocabulary fordescribing the responsibility each player has within a team.Even though it varies from sport to sport, within each sportthese descriptions generalize. The language used is in terms offormations, which is effectively a strategic concept (i.e., dif-ferent teams can use the same formation simultaneously).As a result, we refer to a formation’s generic players usinga set of identity agnostic labels which we denote roles. Aformation is generally shift-invariant and allows for non-rigiddeformations. Therefore, we define each role by its positionrelative to the other roles (i.e., in soccer a left-midfielderplays in-front of the left-back and to the left of the center-midfielder). Each role within a formation is unique (i.e., notwo players within the same formation can have the same roleat the same time), and players can swap roles throughout thematch. Additionally, multiple formations may exist which canbe interpreted as different sets of roles. A role represents anyarbitrary 2D probability density function. Therefore, we canrepresent it non-parametrically by quantizing the field into adiscrete number of cells, or parametrically using a mixtureof 2D Gaussians. We can then represent the formation byconcatenating the features of each role into a single vector.
Pass Foul - Cross CatchDirect FK Drop Save
Pass Foul - Cross CatchAssist Indirect FK Assist Save
Corners Foul - Reception PunchPenalty
Shot on Foul - Reception PunchTarget Throw-in Assist Save
Shot off Offside Reception DivingTarget SaveGoal Yellow Catch Diving
Card SaveOwn Red Catch Drop ofGoal Card Drop Ball
Neutral Running Chance SubstitutionClear Save with Ball
Block Drop Pass Hold ofKick Save Ball
Clearance Neutral Player ClearanceUncontrolled Clearance Out
TABLE II. LIST OF MATCH STATISTICS USED TO DESCRIBE TEAMBEHAVIOR.
Role is a dynamic label, meaning that a player can fulfill manyroles during the game (e.g., a player may switch between left-winger and center-midfielder). However, each role needs to beassigned to a player in every frame so two players can not bein the same role at the same time.
As a formation basically assigns an area or space toeach player at every frame, this problem can be framedas a minimum entropy data partitioning problem [24], [25].Bialkowski et al. [3] show the full derivation, but in practiceit is similar to k-means clustering with the caveat of insteadof assigning each data point to its closest cluster, we solve alinear assignment problem between identities and roles usingthe Hungarian algorithm [26] at each frame. The process isshown in Fig 2. Using this procedure, the resulting formationof each team in every half we analyzed is shown in Fig 3. Inthe next section, we compare the formation descriptor to othermatch factors.
IV. PREDICTING TEAM IDENTITY
To determine if teams had a distinct playing style, weconducted a series of team identity experiments. The challengewas, given only player tracking data and ball events, canwe predict the identity of each team? To do this, we needdescriptors of team behaviors during a match. For this paper,we generated three types of match descriptors: 1) matchstatistics, 2) ball occupancy, and 3) team formation.
A B C D E
F G H I J
K L M N O
P Q R S T
Fig. 3. Example of our formation descriptors for each team. The colors represent different roles. For visualization purposes we have just plotted the centroidfor each role for each match.
A. Match Descriptors
Match Statistics: During a match, various statistics thatcapture team and individual behavior are annotated. Table IIshows the list of statistics which we used in this paper. Whilethe number of these match statistic is quite large, the majorityof them are quite sparse with only a couple of these eventslabeled per match. In reporting of a match, only a half-dozen ofthe most important match statistics are normally documented(i.e., goals, shots on target, shots off target, passes, corners,yellow and red-cards).
Ball Occupancy: Associated with the match statis-tics/events are the time and location for each occurrence.To form a representation of this information, we adoptedthe approach used in [17], [19] which involves estimatingthe continuous ball trajectory at each time-stamp by linearlyinterpolating between events, as well as which team hadpossession (ignoring stoppages). We then broke the field intoa 10 × 8 spatial grid and calculated the ball occupancy ofeach of these grids for each team (i.e. how often the teamwas in possession of the ball in this location over the match).All teams were normalized to attack from left to right. Avisualization of a resulting ball occupancy example is shownin Fig. 4.
Formation Descriptor: For each match half, we foundthe formation descriptor F∗ by using the method described inSection III. This gave an M×N matrix where M refers to thenumber of cells in the field and N is the number of roles (setto 10, as we omitted the goal-keeper as well as games which
M185 T1 − Occupancy map
Fig. 4. Example ball occupancy map over a match half for a team attackingleft to right. This example shows dominance of ball possession on the leftside of the field which may be indicative of the team’s playing style.
had a player sent off). A depiction of the formation descriptorsfor each team for all matches is shown in Fig. 3. For clarity ofpresentation, we have only plotted the centroid of each role foreach match, with each team attacking from left-to-right. Eachdifferent color marker corresponds to a different role for thatteam. It can be seen from the plot that teams are rather rigidin the way they play across a season which suggests that thisis a useful feature in discriminating between different teams.Another interesting point is, as teams vary little in terms ofplaying style throughout the season, this could be used as apowerful prior for preparing against an opposition in upcomingmatches.
Confusion matrix 20−NN, using LDA (CCR = 17.13%)
601801312222200006001911000
00070000000067600006
120120712667000000196000
02561370011201670180006760
0600130661307067660000
601270471107000600061306
2560013617110000606611006
19000060600706001200019
619122006003350660110287126
0250137660011760296007180
0060706605701801706060
0000000003204401411007290
121224202061111011200606120201212
0000000605060360000120
6600700675401200110110012
00120000000000002500012
00007000135700011011706
600700600006000063306
000700000501267000060
000700111100061201100006
ARSE
N
ASTO
NCH
ELS
EVER
T
FULA
MLI
VER
MAN
CI
MAN
UDNE
WCA
NORW
I
QUE
ENRE
ADI
SHAM
PST
OKE
SUND
E
SWAN
STO
TTE
WBR
OM
WEH
AMW
IGAN
ARSENASTONCHELSEVERTFULAMLIVER
MANCIMANUDNEWCANORWIQUEENREADI
SHAMPSTOKESUNDESWANSTOTTE
WBROMWEHAMWIGAN
0
10
20
30
40
50
60
70
80
90
100Confusion matrix 20−NN, using LDA (CCR = 19.51%)
3100700111700006011060012
0191207061113520060666766
12600136110205712180600060
00027066070012006001306
0600206000570029012110120
061213029111170060761261306
1266071822675701206120706
251224770628130766002501306
0612700007000601106060
0607006002671260600766
066006116700000006000
00613060001672500600766
01212776611016130014606060
0000700000000216000120
6126076067110612017607619
60007600011066001901306
0000130067013618146050066
0061300000000006001300
600006000076670000240
000070607070070660612
ARSE
N
ASTO
NCH
ELS
EVER
T
FULA
MLI
VER
MAN
CI
MAN
UDNE
WCA
NORW
I
QUE
ENRE
ADI
SHAM
PST
OKE
SUND
E
SWAN
STO
TTE
WBR
OM
WEH
AMW
IGAN
ARSENASTONCHELSEVERTFULAMLIVER
MANCIMANUDNEWCANORWIQUEENREADI
SHAMPSTOKESUNDESWANSTOTTE
WBROMWEHAMWIGAN
0
10
20
30
40
50
60
70
80
90
100Confusion matrix 20−NN, using LDA (CCR = 67.32%)
8166076067000000011000
0380000001300000606000
0065000060500014060006
000737000007000000760
6001373000000060000000
0000094600000600011000
01207008360000000000012
00240706677001260060700
00000006400760000671212
000000007950000600006
0120000000053000600000
0600000013074460606000
0000000670005900061306
0000000000012686000060
0000000000200007200000
000000000000000880000
1260000660076006056766
060070000000600006000
0060000070012600000710
0120700000006000000050
ARSE
N
ASTO
NCH
ELS
EVER
T
FULA
MLI
VER
MAN
CI
MAN
UDNE
WCA
NORW
I
QUE
ENRE
ADI
SHAM
PST
OKE
SUND
E
SWAN
STO
TTE
WBR
OM
WEH
AMW
IGAN
ARSENASTONCHELSEVERTFULAMLIVER
MANCIMANUDNEWCANORWIQUEENREADI
SHAMPSTOKESUNDESWANSTOTTE
WBROMWEHAMWIGAN
0
10
20
30
40
50
60
70
80
90
100Confusion matrix 20−NN, using LDA (CCR = 70.38%)
880000000700060000000
044000000700007600000
6065200060750000006766
060730000000000000000
000080000000000006000
006009406130000006110012
01212000670000060066000
600000683000000000706
01960000040027060600706
066000000950000666700
000000000053000000000
000000000007500600000
0600700070067100061300
000000000000086000060
000000607076077200000
000000000000000810000
060013611117000120006101812
006000607070000005300
000000000076006007710
000700000006000000056
ARSE
N
ASTO
NCH
ELS
EVER
T
FULA
MLI
VER
MAN
CI
MAN
UDNE
WCA
NORW
I
QUE
ENRE
ADI
SHAM
PST
OKE
SUND
E
SWAN
STO
TTE
WBR
OM
WEH
AMW
IGAN
ARSENASTONCHELSEVERTFULAMLIVER
MANCIMANUDNEWCANORWIQUEENREADI
SHAMPSTOKESUNDESWANSTOTTE
WBROMWEHAMWIGAN
0
10
20
30
40
50
60
70
80
90
100
A B TC D E F G H I J K L NM PO Q SR A B TC D E F G H I J K L NM PO Q SR A B TC D E F G H I J K L NM PO Q SR A B TC D E F G H I J K L NM PO Q SR
AB
T
CDEFGHIJKL
NM
PO
Q
SR
Act
ual T
eam
Predicted Team Predicted Team Predicted Team Predicted Team(a) (b) (c) (d)
Fig. 5. Team identity results for the various descriptors presented as confusion matrices, showing the percentage of agreement, with the actual team on thevertical axis and the predicted team on the horizontal (normalized as percentages): (a) match statistics (17.13%), (b) ball occupancy (19.51%), (c) formationdescriptor (67.32%) and (d) fused all descriptors (70.38%).
Get Match Descriptor
Scale Data LDA Predict Team Identity
Figure 6: Example of how our approach works. NB: for visualization purposes we estimate the occupancymaps via covariances for each role which are depicted by ellipses.
used the team identity as the class labels (i.e., C = 20).We learn a W for each feature set and then multiply thefeatures by W to yield a C � 1 feature vector. To predictthe identity label of the teams in the test match, we usea k-nearest-neighbor classifier (k = 20) using the euclideannorm as our distance metric.
XXscale
WLDA
WTLDA
arg maxW
Tr(W⌃bW
W⌃wW) (18)
Maybe put something in here about clustering on style...
5. PREDICTING FUTURE PERFORMANCES
5.1 Predicting Team BehaviorBut generally, these methods are essentially equivalent to
non-negative matrix factorization, kernel k-means and dis-criminative k-means. I think we have to do all three andshow we can similar performance (these coe�cients will es-sentially be our style vector).
This can be seen as discriminative clustering, which is sim-ilar to kernel k-means and is similar to non-negative matrixfactorization.
In the previous section, given we had the ball and player
0
20
40
60
80
Match Stats Ball Occ Formation Combined
Figure 8: Results of Team ID results.
tracking data, we wanted to predict the team identity. Inthis section, we want to do the reverse - given we just havethe identity of the two teams playing, can we predict how thegame will be played by estimating what the match featureswill be.
We use K-NN regression by using the style prior as theinput. From the previous section we have the weights foreach team. Our input is a joint representation of the style
Figure 6: Example of how our approach works. NB: for visualization purposes we estimate the occupancymaps via covariances for each role which are depicted by ellipses.
used the team identity as the class labels (i.e., C = 20).We learn a W for each feature set and then multiply thefeatures by W to yield a C � 1 feature vector. To predictthe identity label of the teams in the test match, we usea k-nearest-neighbor classifier (k = 20) using the euclideannorm as our distance metric.
XXscale
WLDA
WTLDA
arg maxW
Tr(W⌃bW
W⌃wW) (18)
Maybe put something in here about clustering on style...
5. PREDICTING FUTURE PERFORMANCES
5.1 Predicting Team BehaviorBut generally, these methods are essentially equivalent to
non-negative matrix factorization, kernel k-means and dis-criminative k-means. I think we have to do all three andshow we can similar performance (these coe�cients will es-sentially be our style vector).
This can be seen as discriminative clustering, which is sim-ilar to kernel k-means and is similar to non-negative matrixfactorization.
In the previous section, given we had the ball and player
0
20
40
60
80
Match Stats Ball Occ Formation Combined
Figure 8: Results of Team ID results.
tracking data, we wanted to predict the team identity. Inthis section, we want to do the reverse - given we just havethe identity of the two teams playing, can we predict how thegame will be played by estimating what the match featureswill be.
We use K-NN regression by using the style prior as theinput. From the previous section we have the weights foreach team. Our input is a joint representation of the style
Figure 6: Example of how our approach works. NB: for visualization purposes we estimate the occupancymaps via covariances for each role which are depicted by ellipses.
used the team identity as the class labels (i.e., C = 20).We learn a W for each feature set and then multiply thefeatures by W to yield a C � 1 feature vector. To predictthe identity label of the teams in the test match, we usea k-nearest-neighbor classifier (k = 20) using the euclideannorm as our distance metric.
XXscale
WLDA
WTLDA
WTLDAXscale
arg maxW
Tr(W⌃bW
W⌃wW) (18)
Maybe put something in here about clustering on style...
5. PREDICTING FUTURE PERFORMANCES
5.1 Predicting Team BehaviorBut generally, these methods are essentially equivalent to
non-negative matrix factorization, kernel k-means and dis-criminative k-means. I think we have to do all three andshow we can similar performance (these coe�cients will es-sentially be our style vector).
This can be seen as discriminative clustering, which is sim-ilar to kernel k-means and is similar to non-negative matrixfactorization.
0
20
40
60
80
Match Stats Ball Occ Formation Combined
Figure 8: Results of Team ID results.
In the previous section, given we had the ball and playertracking data, we wanted to predict the team identity. Inthis section, we want to do the reverse - given we just havethe identity of the two teams playing, can we predict how thegame will be played by estimating what the match featureswill be.
We use K-NN regression by using the style prior as theinput. From the previous section we have the weights for
Learn LDA Transform
Figure 6: Example of how our approach works. NB: for visualization purposes we estimate the occupancymaps via covariances for each role which are depicted by ellipses.
used the team identity as the class labels (i.e., C = 20).We learn a W for each feature set and then multiply thefeatures by W to yield a C � 1 feature vector. To predictthe identity label of the teams in the test match, we usea k-nearest-neighbor classifier (k = 20) using the euclideannorm as our distance metric.
XXscale
WLDA
WTLDA
WTLDAXscale
W = arg maxW
Tr(W⌃bW
W⌃wW) (18)
Maybe put something in here about clustering on style...
5. PREDICTING FUTURE PERFORMANCES
5.1 Predicting Team BehaviorBut generally, these methods are essentially equivalent to
non-negative matrix factorization, kernel k-means and dis-criminative k-means. I think we have to do all three andshow we can similar performance (these coe�cients will es-sentially be our style vector).
This can be seen as discriminative clustering, which is sim-ilar to kernel k-means and is similar to non-negative matrixfactorization.
0
20
40
60
80
Match Stats Ball Occ Formation Combined
Figure 8: Results of Team ID results.
In the previous section, given we had the ball and playertracking data, we wanted to predict the team identity. Inthis section, we want to do the reverse - given we just havethe identity of the two teams playing, can we predict how thegame will be played by estimating what the match featureswill be.
We use K-NN regression by using the style prior as theinput. From the previous section we have the weights for
Train
Fig. 6. Given a match descriptor, we first scale the data and then multiply it byWT which is found using LDA to yield a discriminative feature vector. TheLDA matrix is learnt using the team identity labels and their match descriptorsin the training set. Team identity is predicted using k-NN.
B. Experiments
The team identity experiments were performed usinga “leave-one-match-out” cross-validation strategy where onematch was left out to test against, and the remaining matcheswere used as the train set. A block diagram in Fig. 6 de-scribes the process. Firstly, we generated the three descriptorsdescribed above and scaled the features. To obtain a compactbut discriminative representation, we performed linear discrim-inant analysis (LDA) by learning the transformation matrixW from the training set and used the team identity as theclass labels (i.e., C = 20). We learnt a W for each descriptorand then multiplied the features by WT to yield a lowerdimensionality discriminant feature vector of dimensionalityC − 1. To predict the identity label of the teams in the testmatch, we used a k-nearest-neighbor classifier (k = 20) usingthe Euclidean norm as the distance metric.
The results for the various descriptors are shown in Fig. 5.In the first experiment, (Fig. 5(a)) we can see that usingonly match statistics is a poor indication of team identitywith an overall accuracy of 17% (chance is 5%). This resultmakes sense as the match statistics only contain coarse eventinformation without any spatial or temporal information aboutthe ball or the players. Using the ball occupancy only gavemarginally improved performance over the match statisticswith an accuracy of 19% (Fig. 5(b)). This is well below the33% which was obtained in the previous works [17], [19]. Apossible explanation of the performance difference could bedue to the coarse estimation of the possession strings and theball occupancy maps from the event data.
The most impressive performance by far is the formationdescriptor which obtains over 67% accuracy, which clearlyshows that teams have a true underlying signal which can beencapsulated in the way the team moves in formation over
Match Stats Ball Occ. Formation Combined0
20
40
60
80
17.13% 19.51%
67.32% 70.38%
Team
IDPr
edic
tion
Acc
urac
y(%
)
Fig. 7. Comparison of team identity prediction accuracy for the three differentmatch descriptors, as well as the combined performance.
time (Fig. 5(c)). We also fused together these descriptors byconcatenating all the scaled features, and performing LDA onthe combined features. This approach improved the overallperformance to over 70% which shows there is complimentaryinformation within the other descriptors. A bar-graph compar-ing the overall performance for each descriptor is given inFig. 7.
V. ANALYZING TEAM BEHAVIORS
In this section we explore how we can learn and representthe characteristic style of teams, and use this for analyzingteam behaviors in prediction and anomaly detection tasks.
A. Team Style
Team style is a very subjective and high-level attribute tolabel, especially in continuous sports like soccer. This is in partdue to the dynamic and low-scoring nature of such sports, asit is hard to segment the game into discrete parts and assign alabel when style encompasses all aspects of play. Due to theglobal nature of style, one way to quantify a team’s style isvia a linear combination of prior behavior styles.
Given a training set of team behavior descriptors, we candiscover a discrete set of styles using k-means clustering.For evaluation, we exclude the last two rounds of the seasonfor testing, and use the remaining games to train the stylemodels. We first project the match features into a lowerdimensional, discriminative space using LDA, as in the teamidentity experiments (Fig. 6), and then cluster similar examples
AB
T
CDEFGHIJKL
NM
PO
Q
SR
Act
ual T
eam
030030500044011500511900
271312532293110029002620402
11102100224242904232011611
00000000070281290100310
000000000000000010028
1 2 3 4 5
ARSENASTONCHELSEVERTFULAMLIVER
MANCIMANUDNEWCANORWIQUEENREADI
SHAMPSTOKESUNDESWANSTOTTE
WBROMWEHAMWIGAN
0
5
10
15
20
25
30
1 2 3 4 5
031001000001000000000
00231401311012110270000
009003270000000020000
26017040529101140022001
00001500015342001282900001
000000000002902310000
2031701616060300031011
001000005010000002900
000000001000000000310
010000000000000000028
1 2 3 4 5 6 7 8 9 10
ARSENASTONCHELSEVERTFULAMLIVER
MANCIMANUDNEWCANORWIQUEENREADI
SHAMPSTOKESUNDESWANSTOTTE
WBROMWEHAMWIGAN
0
5
10
15
20
25
30
1 2 3 4 5 6 7 8 9 10
029000000000000000000
003200005000200000000
000150000000100000000
000150000000000000000
000029000010000001002
000003200000000000000
000000290000000000000
2500000026000000001000
010000002704011001011
010020000330000000000
000000000024000100000
200200010000280001000
000000000100030000000
000000000000003100000
000000000000000320000
100000011010000029011
000000001000000002900
0000000000028000000300
010000000000000000015
000000000000000000012
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
ARSENASTONCHELSEVERTFULAMLIVER
MANCIMANUDNEWCANORWIQUEENREADI
SHAMPSTOKESUNDESWANSTOTTE
WBROMWEHAMWIGAN
0
5
10
15
20
25
30
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 1920Style Cluster Style Cluster Style Cluster
(a) (b) (c)
Fig. 8. Results for clustering descriptors of each match half when we set the number of style clusters to: (a) 5, (b) 10, and (c) 20. These can be used as astyle prior for predicting the results of future matches.
Fig. 10. Prediction of formation using k-NN regression. (a) all training examples, (b) retrieved examples according to style prior, (c) the predicted formation(= mean(retrieved examples)), (d) the actual formation.
Team’s style ordered by date(Note, some values are missing so same column does not correspond to same round)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41
ARSENASTONCHELSEVERTFULAMLIVER
MANCIMANUDNEWCANORWIQUEENREADI
SHAMPSTOKESUNDESWANSTOTTE
WBROMWEHAMWIGAN
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
5AB
T
CDEFGHIJKL
NM
PO
Q
SR
Act
ual T
eam
Team’s style ordered by date(Note, some values are missing so same column does not correspond to same round)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41
ARSENASTONCHELSEVERTFULAMLIVER
MANCIMANUDNEWCANORWIQUEENREADI
SHAMPSTOKESUNDESWANSTOTTE
WBROMWEHAMWIGAN
0
0.5
1
1.5
2
2.5
3
3.5
4
4.5
51 2 3 4 5Style Cluster:
Team
ID
Fig. 9. The variation in team style for each team across a season can easilybe seen when to 5 styles.
in this space. The style clustering results for k = 5, 10 and 20,are shown in Fig. 8.
Observing Fig. 8, there is some overlap in styles betweencertain teams, and some teams exhibit multiple styles. Thevariation in style for each team using k = 5 styles, is shown inFig. 9. Team T stands out, being in a style cluster of its own,which could be explained by the distinctly different formation
from all other teams, with 3 defenders at the back (see Fig. 3).Most teams play a single style, while teams E and R vary theirplaying styles more frequently than other teams.
To encapsulate the behavior styles that teams adopt, wedefine the playing style of a team as the normalized weightsfrom the style clustering matrices (e.g. for the 5 style clustersused in Fig. 8(a), the style vector for Team A=[0, 27
28 , 128 , 0, 0],
Team B=[ 3032 ,132 ,
132 , 0, 0], etc.). Modeling teams as a
combination of the styles they play makes intuitive sense, assometimes a team could play a pressing game and on otheroccasions the team may play defensively, so they would beweighted according to these performances. Another team maybe very rigid and play the same style every game - so theweight for that style may be very high. These style vectorscan then be used to assist prediction.
B. Prediction and Anomaly Detection
Previously, given the ball and player tracking data, wepredicted the team identity. In this section, we want to dothe reverse - given we just have the identity of the two teamsplaying, can we predict how the game will be played byestimating what the match features will be?
To predict the most likely features, we use K-NN regressionusing the learnt team style priors as the input, which allowsus to select which of the training matches to regress from for
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
2
4
6
Test match index
Mea
ner
ror
per
role
(m)
Difference Between Predicted and Actual Formation Played
Home TeamAway Team
Fig. 11. Results of comparing the predicted formation to the actual formationplayed for each match in the final two rounds of the season (18 matches).
−200 0 200
−200
−100
0
100
200
−200 0 200
−200
−100
0
100
200
*MANUD* (Home) vs WBROM (Away) (distBetween est & actual = 9.71)
−200 0 200
−200
−100
0
100
200
−200 0 200
−200
−100
0
100
200
−200 0 200
−200
−100
0
100
200
*MANUD* (Home) vs WBROM (Away) (distBetween est & actual = 9.71)
−200 0 200
−200
−100
0
100
200
−200 0 200
−200
−100
0
100
200
−200 0 200
−200
−100
0
100
200
*MANUD* (Home) vs WBROM (Away) (distBetween est & actual = 9.71)
−200 0 200
−200
−100
0
100
200
(a) (b) (c)
Fig. 12. Example of a poor formation estimate (test match 16), which appearsto be due to an anomaly in the team’s behavior. (a) retrieved examples, (b)predicted formation, (c) actual formations
our prediction. That is, for each match in the training set, wecompare the two team styles to the test match’s team stylepriors. We then extract the matches which are most similar interms of team styles, and calculate the mean features to predictthe outcome of the test match. We can then compare thisprediction with the actual result. The procedure, demonstratingformation prediction is shown in Fig. 10.
We performed prediction of team formation on the last tworounds of the season (containing 18 matches) and evaluatedthe results by comparing the predicted formation to the actualformation played as presented in Fig. 11. It can be seen thatmost matches are estimated within 2 m average error per role,while Match 1 and 16 are most poorly estimated. This suggeststhat the teams were not playing their normal formation style inthese matches (i.e. anomalous behavior). The predictions allowus to visualize the most likely formation given prior examplesand when anomalies occur, such as in Fig. 12.
VI. SUMMARY AND FUTURE WORK
In this paper, we first presented a formation descriptorwhich was found by minimizing the entropy of a set of playerroles. Using an entire season of player tracking data, wegenerated the formation descriptor by projecting the set of oc-cupancy maps of each role into a low-dimensional discrimina-tive feature space using linear discriminating analysis (LDA).We showed that this approach characterizes individual teambehavior significantly better (3 times more) than other matchdescriptors which are normally used to describe team behavior.We then conducted a series of analysis and predictions whichshowed the utility of our approach. In future work, we planto use this descriptor for short-term prediction (i.e., who willthe next pass go to etc.), as well as long-term prediction (i.e.,match result).
Acknowledgement: The QUT portion of this research was supported by theQld Govt’s Dept. of Employment, Economic Development & Innovation.
REFERENCES
[1] Prozone, www.prozonesports.com.[2] STATS SportsVU, www.sportvu.com.[3] A. Bialkowski, P. Lucey, P. Carr, Y. Yue, S. Sridharan, and I. Matthews,
“Large-Scale Analysis of Soccer Matches using Spatiotemporal Track-ing Data,” in ICDM, 2014.
[4] K. Goldsberry, “CourtVision: New Visual and Spatial Analytics for theNBA,” in MIT SSAC, 2012.
[5] R. Masheswaran, Y. Chang, A. Henehan, and S. Danesis, “Destructingthe Rebound with Optical Tracking Data,” in MIT SSAC, 2012.
[6] R. Masheswaran, Y. Chang, J. Su, S. Kwok, T. Levy, A. Wexler, andN. Hollingsworth, “The Three Dimensions of Rebounding,” in MITSSAC, 2014.
[7] J. Wiens, G. Balakrishnan, J. Brooks, and J. Guttag, “To Crash or Notto Crash: A quantitative look at the relationship between the offensiverebounding and transition defense in the NBA,” in MIT SSAC, 2013.
[8] P. Lucey, A. Bialkowski, P. Carr, Y. Yue, and I. Matthews, “How to Getan Open Shot: Analyzing Team Movement in Basketball using TrackingData,” in MIT SSAC, 2014.
[9] A. Bocskocsky, J. Ezekowitz, and C. Stein, “The Hot Hand: A NewApproach to an Old “Fallacy”,” in MIT SSAC, 2014.
[10] A. Miller, L. Bornn, R. Adams, and K. Goldsberry, “Factorized PointProcess Intensities: A Spatial Analysis of Professional Basketball,” inICML, 2014.
[11] D. Cervone, A. D’Amour, L. Bornn, and K. Goldsberry, “POINTWISE:Predicting Points and Valuing Decisions in Real Time with NBA OpticalTracking Data,” in MIT SSAC, 2014.
[12] P. Carr, M. Mistry, and I. Matthews, “Hybrid Robotic/Virtual Pan-Tilt-Zoom Cameras for Autonomous Event Recording,” in ACM Multimedia,2013.
[13] X. Wei, P. Lucey, S. Morgan, and S. Sridharan, “Sweet-Spot: UsingSpatiotemporal Data to Discover and Predict Shots in Tennis,” in MITSSAC, 2013.
[14] ——, “Predicting Shot Locations in Tennis using Spatiotemporal Data,”in DICTA, 2013.
[15] G. Ganeshapillai and J. Guttag, “A Data-Driven Method for In-GameDecision Making in MLB,” in MIT SSAC, 2014.
[16] S. Sinha, C. Dyer, K. Gimpel, and N. Smith, “Predicting the NFL UsingTwitter,” in ECML Workshop on Machine Learning and Data Miningfor Sports Analytics, 2013.
[17] P. Lucey, A. Bialkowski, P. Carr, E. Foote, and I. Matthews, “Character-izing Multi-Agent Team Behavior from Partial Team Tracings: Evidencefrom the English Premier League,” in AAAI, 2012.
[18] Opta Sports, www.optasports.com.[19] P. Lucey, D. Oliver, P. Carr, J. Roth, and I. Matthews, “Assessing team
strategy using spatiotemporal data,” in ACM SIGKDD, 2013.[20] A. Bialkowski, P. Lucey, P. Carr, Y. Yue, and I. Matthews, “Win at
home and draw away: Automatic formation analysis highlighting thedifferences in home and away team behaviors,” in MIT SSAC, 2014.
[21] J. Tenenbaum and W. Freeman, “Separating Style and Content withBilinear Models,” Neural Computation, vol. 12, no. 6, pp. 1247–1283,2000.
[22] C. Doersch, S. Singh, A. Gupta, J. Sivic, and A. Efros, “What MakesParis Look Like Paris?” ACM Transactions on Graphics (SIGGRAPH),vol. 31, no. 4, 2012.
[23] Y. Lee, A. Efros, and M. Hebert, “Style-Aware Mid-Level Representa-tion for Discovering Visual Connections in Space and Time,” in ICCV,2013.
[24] S. Roberts, R. Everson, and I. Rezek, “Minimum Entropy Data Parti-tioning,” IET, pp. 844–849, 1999.
[25] Y. Lee and S. Choi, “Minimum Entropy, K-Means, Spectral Clustering,”in International Joint Conference on Neural Networks, 2004.
[26] H. W. Kuhn, “The hungarian method for the assignment problem,”Naval Research Logistics Quarterly, vol. 2, no. 1-2, pp. 83–97, 1955.