2

It is a general question on how to learning representation of one entity but the dataset is mixed with a lot of other entities, which their statis are always waiting to be learnt.

The question is best be explained by an example. Let's say, the entity is a football player. How can we assign player stats to them by just learning from the match result.

Dataset: Football match data with startup line of the two teams and the final result score.

Player stats, something like this https://www.fifaindex.com/player/158023/lionel-messi/ but much simpler. We can define a player stats is just 3D vector v = (x , y , z) where x , y , z is between 0 to 1.0. We don't need to know what actually x mean shooting power or anything.

So after learning, It output each player a v vector assigned.

Any idea or some material that you can suggest?

palazzo train
  • 121
  • 3
  • How can you tell between two players who have played together in every match? Coaches may try and keep teams consistent so there's not much information about a single player from a set of game results. What does your data really look like? – Spacedman May 25 '18 at 07:42

1 Answers1

2

For that, you would probably need something more granular than just the score (e.g., in-game analytics, etc.) to help you assess the attributes of each player. Even then, it would be difficult to map the outcome to multiple attributes (i.e., defending skills, offensive skills).

If you only care about a single score of a player (i.e., say, how effective is a player in helping her team to win, etc.), you could potentially look at something like Corsi which essentially measures puck possession (not without its criticism or flaws). There are many variants out there, but it should give you an idea how to leverage in-game analytics to assess a player.

But then, Corsi is merely a calculation from in-game statistic; you will need more than that to call it a machine learning exercise.

The Lyrist
  • 474
  • 3
  • 14