Graph-based Approaches for Analysing Team Interaction on the - - PowerPoint PPT Presentation

Graph-based Approaches for Analysing Team Interaction

• n the Example of Soccer

Markus Brandt and Ulf Brefeld

Knowledge Mining & Assessment TU Darmstadt / DIPF

brefeld@cs.tu-darmstadt.de

Ulf Brefeld brefeld@cs.tu-darmstadt.de

Motivation

๏ Team interaction important for playmaking, team

strategy and thus also for winning

๏ BUT: interaction not covered well by counting-

based approaches and descriptive statistics

๏ Here: use of Page rank to capture team interaction

2

Ulf Brefeld brefeld@cs.tu-darmstadt.de

Representation

๏ Importance of

team interaction

๏ Credits for both

involved players

3

pass-based receiver-based interaction-based

๏ Importance of

successful pass

๏ Credits for

player who passes the ball

๏ Captures

natural flow

๏ Receiver gets

the credits

Ulf Brefeld brefeld@cs.tu-darmstadt.de

Representation

๏ Importance of

team interaction

๏ Credits for both

involved players

4

pass-based receiver-based interaction-based

๏ Importance of

successful pass

๏ Credits for

player who passes the ball

๏ Captures

natural flow

๏ Receiver gets

the credits

Ulf Brefeld brefeld@cs.tu-darmstadt.de

Representation

๏ Importance of

team interaction

๏ Credits for both

involved players

5

pass-based receiver-based interaction-based

๏ Importance of

successful pass

๏ Credits for

player who passes the ball

๏ Captures

natural flow

๏ Receiver gets

the credits

Ulf Brefeld brefeld@cs.tu-darmstadt.de

Chain Score

๏ Average length of chains a player p is involved in:

6

Page Rank

๏ Normalised sum of page ranks of nodes pointing to

p given by

Ulf Brefeld brefeld@cs.tu-darmstadt.de

Empirical Results

๏ Bundesliga seasons 2009/10 - 2013/14 ๏ 1530 games, >250000 pass chains ๏ Player ranking ๏ Prediction of winning team

7

distribution of chain lengths

Ulf Brefeld brefeld@cs.tu-darmstadt.de

Top-ranked Players

๏ Chain score ranks players according to #chains ๏ Page rank captures team interaction + playmaking

8

chain score page rank (pass) page rank (receiver) page rank (interaction)

Ulf Brefeld brefeld@cs.tu-darmstadt.de

Preliminary Results

๏ Baselines: home team: 47.39%, better team 52.94% ๏ 2 features (home and away team score) ๏ Chain score worst, completed passes and page rank

perform similarly

9

chain score pass interact pass receive interact pass receive interact page rank

train: 2009/10 - 2012/13, test 2013/14

receive #completed passes

Ulf Brefeld brefeld@cs.tu-darmstadt.de

Conclusion

๏ Page rank captures team interaction and playmaking

better than simple frequency-based approaches

๏ Representation depends on application at-hand (e.g.,

pass, receiver, interaction-based)

๏ Preliminary results for predicting winning teams

show that counting-based baselines (#completed passes) perform similarly

10