Identifying Team Style in Soccer using Formations from ... · Identifying Team Style in Soccer...
Transcript of Identifying Team Style in Soccer using Formations from ... · Identifying Team Style in Soccer...
Identifying Team Style in Soccer using Formationsfrom 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 show the utility ofour approach using a entire season of Prozone player trackingdata 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, an expert human is able to do this but this isvery labor intensive and is inherently subjective. Having amethod which can quantify these behaviors should be possiblewith the prevalence of spatiotemporal tracking data of playerand ball movement being captured in most professional sports(e.g., [1], [2]). However, this task is challenging due to thecomplexities dealing with adversarial multi-agent trajectorydata. A major issue centers on the alignment of individualplayer trajectories within a team setting which is a sourceof noise. In this paper, we align the data based on a role-based method which is learnt directly from data [3]. We showthat using this approach, semantically meaningful team-basedstrategic features can be obtained which are highly predictiveof their identity. We compare this descriptor to other featuressuch as match statistics (e.g., shots, passes, fouls) and ballmovement and show that the formation descriptor is far moresuperior in discriminating unique team characteristics (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 teamsachieved open three-point shots. Bocskocksy et al. [9] re-investigated the hot-hand theory. Miller et al., [10] analyzed
A B C D E F G H I J K L M N O P Q R S T
Team ID
Statistics Ball occupancy FormationShots (on goal) 12(4)
Fouls 11
Corner kicks 8
Offsides 4
Time of possession 62%
Yellow cards 1
Red cards 0
Saves 3
Game716, T1, GT Label = 4−1−4−1
12
34
56
7
8
9
10{WEHAMAverage overall
Fig. 1. In this paper, based solely on (left) match statistics, (middle) ballmovement patterns, and (right) formation descriptor - we can predict with highaccuracy the identity of a team soccer. We show the formation descriptor isthe best discriminator of team style.
the 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 in thatlocation for automatic sport broadcasting purposes. In tennis,Wei et al. [13], [14] used Hawk-Eye data to predict the typeand location of the next shot. Ganeshapillai and Guttag [15]used SVMs to predict pitching in baseball while Sinha etal. [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 thehome 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 mean-normalizedcovariances of player positions which again highlights the overlap. (c) Using our iterative approach (which is very similar to k-means with the constraint thatat every frame each detection requires a unique role), a role label is assigned to each player at the frame-level, we 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 consisted of 20 teamswho played home and away, totaling 38 games for each teamor 380 games overall. Five of these games were omitted forvarious reasons. 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., insoccer a left-midfielder playsin-front of the left-back and to the left of the center-midfielder).Each role within a formation is unique (i.e., no two playerswithin the same formation can have the same role at thesame time), and players can swap roles throughout the match.Additionally, multiple formations may exist which can beinterpreted 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 be fulfillmany roles during the game (e.g., left-winger switches to theright-wing and is not characterize as the left-winger becausehe/she started there). However, each role needs to be assignedto a player in every frame (i.e., two players can not be in thesame 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]. The process is shown in Fig 2.The formation of every team in every half we analyzed inshown in Fig 3. We compare the formation descriptor to othermatch factors in the next section.
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.
M185 T1 − Occupancy map
Fig. 4. Example of ball occupancy map of a team from a match (attackingleft to right).
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 are quite large, the majorityof them are quite sparse with only a couple of these eventslabelled per match. In reporting in the match, only a half-dozen of the most important match statistics are normallydocumented (i.e., goals, shots on target, shots off target, passes,corners, yellow and red-cards).
Ball Occupancy: Associated with the match statistics arethe time and location that the event occurred. To form arepresentation of this information, we adopted the approachedused in [17], [19] which consists of estimating the continuous
ball trajectory at each time-stamp as well as which team hadpossession (we ignore stoppages). We then broke the field intoa 10×8 spatial grid and calculated the ball occupancy of eachof these grids for each team (i.e. how often the team wasin possession of the ball in this location over the match). Avisualization of the resulting occupancy is shown in Figure 4.
Formation Descriptor: For each match half, we foundthe formation descriptor F∗ by using the method described inSection III. This gave a M × N matrix where M refers tothe number of cells in the field and N is the number of roleswhich was 10 (we omitted the goal-keeper as well as gameswhich had a player sent off). A depiction of the formationdescriptors for each team for all matches are shown in Figure 3.For clarity of presentation, we have only plotted the centroidof each role for each match with each team attacking from left-to-right. Each different color marker corresponds to a differentrole for that team. It can be seen from the plot that teams arerather rigid in the way they play across a season which suggestthat this is a useful feature in discriminating between differentteams. Another interesting point, as teams vary little in termsof playing style throughout the season, this could be used asa powerful prior in opposition teams preparing for upcomingmatches.
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 matches
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
100AB
T
CDEFGHIJKL
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 A B TC D E F G H I J K L 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: (a) match statistics, (b) ball occupancy, (c) formation descriptor and (d) fused all descriptors.
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.
were used as our train set. The results are shown in Figure 5and the block-diagram shown in Figure 6 describes the process.Firstly, we generate the descriptors described above and thenscale the features. To obtain a compact but discriminativerepresentation, we perform linear discriminant analysis (LDA)by learning the transformation matrix W from the trainingset where we used the team identity as the class labels(i.e., C = 20). We learn a W for each descriptor and thenmultiply the features by W to yield a C−1 feature vector. Topredict the identity label of the teams in the test match, we usea k-nearest-neighbor classifier (k = 20) using the Euclideannorm as our distance metric.
The results for the various descriptors are shown in Fig-ure 5. In the first experiment, (Figure 5(a)) we can see thatusing the match statistics are quite low with an overall accuracyof 17% (chance is 5%). This result makes sense as the matchstatistics only contain coarse event information without anyspatial or temporal information about the ball or the players.Using the ball occupancy only gave marginally improvedperformance over the match statistics with an accuracy of 19%(Figure 5(b)). This is well below the 33% which was obtainedin the previous works [17], [19]. A possible explanation of theperformance difference could be due to the coarse estimationof the possession strings and the ball occupancy maps fromthe 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 time(Figure 5(c)). We also fused together these descriptors by late-integration method where the number of k-nearest neighborsfor each stream was dependent on how reliable the streamwas (i.e., the formation descriptor received the most neigh-
0
20
40
60
80
Match Stats Ball Occ Formation Combined
Team
Iden
tity
Pred
ictio
n A
ccur
acy
(%)
Fig. 7. Comparison of the accuracy of predicting team identity based on thedifferent descriptors.
bors/votes). This approached improved the overall performanceto over 70% which shows there is complimentary informationwithin the other descriptors. A bar-graph comparing the overallperformance for each descriptor is given in Figure 7.
V. ANALYZING TEAM BEHAVIORS
In this section we explore how we can learn and representthe characteristic style of teams, and use this for analysingteam behaviours 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 behaviour styles.
Given a training set of team behaviorr descriptors, wecan discover 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 team
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 1920
(a) (b) (c)
Fig. 8. Shows the clustering results based on style when we set the number of styles to: (a) 5, (b) 10, and (c) 20. These can be used as a style prior forpredicting 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
5
Fig. 9. Shows the variation in style each team has when we cluster to only5 styles.
identity experiments (Fig. 6), and then cluster similar examplesin this space. The style clustering results for k = 5, 10 and 20,are shown in Figure 8.
Observing Figure 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 shownin Figure 9. Team T stands out, being in a style cluster ofits own, which could be explained by the distinctly differentformation from all other teams, with 3 defenders at the back
(see Fig. 3). Most teams play a single style, while teams E andR vary their playing styles more frequently than other teams.
To encapsulate the behaviour styles that teams adopt,we define the playing style of a team as the normalisedweights from the style clustering matrices (e.g. for the5 style clusters used in Figure 8(a), the style vector forTeam 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 makesintuitive sense, as sometimes a team could play a pressinggame and on other occasions the team may play defensively,so they would be weighted according to these performances.Another team may be very rigid and play the same style everygame - so the weight for that game may be very high. Thesestyle vectors can 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 forour prediction. That is, for each match in the training set, we
0 2 4 6 8 10 12 14 16 181
1.5
2
2.5
3
3.5
4
4.5
5
5.5
Mea
n er
ror
for
each
rol
e (m
)
Test match index
Difference between formation estimated & actual
Fig. 11. Evaluation of formation prediction results (prediction vs actual, Red= home team, Blue = away team)
−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
Fig. 12. Example of a poor formation estimate, which appears to be dueto an anomaly in the team’s behaviour. (a) retrieved examples, (b) predictedformation, (c) actual formation
compare 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 Figure 10.
We performed prediction of team formation on the last tworounds of the season (containing 18 matches). We evaluatedthe results by comparing the predicted formation to the actualformation played, presented in Figure 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 stylein these matches (i.e. anomalous behaviour). The predictionsallow us to visualise the most likely formation given priorexamples and when anomalies occur, such as in Figure 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).
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 MITSSAC, 2012.
[5] R. Masheswaran, Y. Chang, A. Henehan, and S. Danesis, “Destruct-ing the Rebound with Optical Tracking Data,” in MIT Sloan SportsAnalytics Conference, 2012.
[6] R. Masheswaran, Y. Chang, J. Su, S. Kwok, T. Levy, A. Wexler, andN. Hollingsworth, “The Three Dimensions of Rebounding,” in MITSloan Sports Analytics Conference, 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 MITSSAC, 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 MITSSAC, 2014.
[9] A. Bocskocsky, J. Ezekowitz, and C. Stein, “The Hot Hand: A New Ap-proach to an Old “Fallacy”,” in MIT Sloan Sports Analytics Conference,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 Sloan Sports Analytics Conference, 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 MITSloan Sports Analytics Conference, 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 Sloan Sports Analytics Conference,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 Sloan SportsAnalytics Conference, 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.