8

I'm trying to analyze a dataset that originates from sensors located near players' shoes in a match (http://www.orgs.ttu.edu/debs2013/index.php?goto=cfchallengedetails).

I decided to look at clustering to identify:

  1. Similar trajectories of players in the match by using the TRACLUS clustering algorithm

  2. Similar players by counting some characteristics such as unsuccessful passages, unsuccessful crosses, shots and tackles. I thought to use DBSCAN to cluster them.

  3. Group players that pass the ball to each other more often. How can I cluster them?

Can I exploit something else from this type of dataset? Is there any other characteristic which I can use in point 2?

Glen_b
  • 282,281
denadai2
  • 89
  • 2

1 Answers1

1

There are 2 questions there (1st point is not a question). All answers are below.

Q1: How can you cluster players that pass the ball to each other more often?

In my view this is a loaded task that is better broken into the following:

  • Identify whether a player is passing a ball. You have to look at the distribution of sensory data that is often associated with actions related to passing balls. Many ways to do this. Once fancy way could be to replicate this emperically-collected dataset in a 3D game that you load players with similar sensors. The nice thing about the game is that you can identify the target variables that you wish to predict (i.e. you know if they are passing the ball). This way, using the game, you can correlate distribution of sensory data to the targeted variables, ultimately generated a labelled set of samples. Finally, you apply a domain-adaptation step by which your 3D game model is transformed to the domain emperically-collecfted dataset (so you can run it there with less error than without the domain adaptation step).
  • Identify whether a player is receiving a ball. Similar to point above but for the distribution of sensory data upon the receipt of balls.
  • Identifying linked passes and receives. This is relatively trivial: two players pass balls to each others iff a receive happens after a pass. To reduce noise, you may wish to add additional constraints to this assumption to ensure that accidental passes are set apart from intentional ones.

Q2: Can I exploit something else from this type of dataset? (so that you expand point 2)

  • Fatigue/stamina/speed as a function of activity and time. This could be possibly easy to estimate by looking at the frequency speed of how the sensors positions/speed are changing.
  • Once you identify the point above, you can estimate other parameters, such as recovery time.
  • Additionally, correlate all above with players relationship with his team. For example, does a player pass balls more often when he is tired? To whom players, or which directions, does he tend to pass balls when he is tired? Does he change his passing targets/directions when he recovers his stamina?
caveman
  • 2,701
  • 2
  • 21
  • 34