I noticed that many of the public note books ignore the \"position\" in the player_tracking data.
\nThese character data (e.g. QB, CB, DE…) were useful data by using knowledge of the NFL domain.
\nI will share my usage.
The NFL is played by 11 players vs 11 players.
\nEach team has three different units.
These unit have distinct roles, each appearing on the field when needed.
\nNFL is a very organized sport, and players on each unit are considered to have different roles.
\nI assumed that each player plays differently, which affects their danger (contact probability).
\nBecause do you think a front line defender and a kicker would be put in the same danger?
I classified the positions into five groups in terms of player roles.
\nI saw the front line players are blocking the opposing players, and behind player are waiting for the ball to come.
\nSo I assumed that the same unit members would have different roles in the front and in the back.
\nThis figure is a simplified representation of each player's position when the Home team on the left is the offense and the Away team on the right is the defense.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10163579%2Fa4abc2c7505a5f72ea78aad9301e0248%2FNFL-position.001.jpeg?generation=1677860871145422&alt=media\")
The positions are divided into groups as shown in the next figure.
\nThe direction to the opposing team is defined as \"front\".
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F10163579%2F15fc0aedf990a551b03fda291d51fd4b%2FNFL-position.002.jpeg?generation=1677860933592393&alt=media\")
The tracking data in training contained 28 unique characters as positions.
\nThese positions classified into 5 groups with knowledge of the NFL domain.
\nI decided on the following, but I do not think it is correct.
\nSome positions have different roles and names depending on the situation.
I have confirmed model with this data improves performance.
\nAs an example, I show the results applied to this public notebook using XGB.
| Model | \nPrivate | \nPublic | \n
|---|---|---|
| Original | \n0.57012 | \n0.58308 | \n
| Add-position | \n0.58933 | \n0.59609 | \n
I have used this data for other models (with images, with table), all of which have improved score.
\nThis was my first time participating in the competition.
\nThe result was 52nd place, which is a good start.
\nI hope to do my best to win a medal again next time.
\nI love Kaggle!!!
I noticed that many of the public note books ignore the \"position\" in the player_tracking data.
\nThese character data (e.g. QB, CB, DE…) were useful data by using knowledge of the NFL domain.
\nI will share my usage.
The NFL is played by 11 players vs 11 players.
\nEach team has three different units.
These unit have distinct roles, each appearing on the field when needed.
\nNFL is a very organized sport, and players on each unit are considered to have different roles.
\nI assumed that each player plays differently, which affects their danger (contact probability).
\nBecause do you think a front line defender and a kicker would be put in the same danger?
I classified the positions into five groups in terms of player roles.
\nI saw the front line players are blocking the opposing players, and behind player are waiting for the ball to come.
\nSo I assumed that the same unit members would have different roles in the front and in the back.
\nThis figure is a simplified representation of each player's position when the Home team on the left is the offense and the Away team on the right is the defense.
\n
The positions are divided into groups as shown in the next figure.
\nThe direction to the opposing team is defined as \"front\".
The tracking data in training contained 28 unique characters as positions.
\nThese positions classified into 5 groups with knowledge of the NFL domain.
\nI decided on the following, but I do not think it is correct.
\nSome positions have different roles and names depending on the situation.
I have confirmed model with this data improves performance.
\nAs an example, I show the results applied to this public notebook using XGB.
| Model | \nPrivate | \nPublic | \n
|---|---|---|
| Original | \n0.57012 | \n0.58308 | \n
| Add-position | \n0.58933 | \n0.59609 | \n
I have used this data for other models (with images, with table), all of which have improved score.
\nThis was my first time participating in the competition.
\nThe result was 52nd place, which is a good start.
\nI hope to do my best to win a medal again next time.
\nI love Kaggle!!!
I found the position was quite useful feature. The boost to the stronger models was expectedly more modest compared to the XGB public notebook you mentioned.
", "rawMarkdown": "I found the position was quite useful feature. The boost to the stronger models was expectedly more modest compared to the XGB public notebook you mentioned.", "votes": null }, { "id": "2168491", "postDate": "03/04/2023 09:33:33", "content": "Thank you.
\nI also think the impact of this is small on stronger models. I just think the position is data that cannot be ignored. There may be a better use for it.
I found the position was quite useful feature. The boost to the stronger models was expectedly more modest compared to the XGB public notebook you mentioned.
", "votes": null, "replies": [ { "id": 2168491, "author_name": "sunagimon", "author_url": "", "post_date": "03/04/2023 09:33:33", "content": "Thank you.
\nI also think the impact of this is small on stronger models. I just think the position is data that cannot be ignored. There may be a better use for it.