ashwinvaswani / whatif Goto Github PK

Currently we do not have the data for the 2019/20 season of the Champions League as the format of the UEFA website is different for these matches. Please scrape the data for these matches separately, and upload the csv to the repository.

Hi,
I'm a student approaching the world of Neural Networks and I was asked to go through your project, but I can't figure out how to run it.
Could you help me please?

Certain clubs such as Valencia, Atlanta and RB Leibzig have no data in our dataset, but they are present in the data that we will be predicting on. We need to scrape local games from https://datahub.io/collections/football in order to get the context for these clubs. We also need the data for every club even if we have examples of them in our database in order to generate the form vector for the clubs. The scrapers will be different for each league, so it will be better for different people to take up different leagues.

Currently our data contains a lot of columns with NAN or blank values. We would like them to be replaced by the default values. (So 0 for numeric columns).

Also for the 2014/15 season we have some missing details which can be manually filled in. Where possible please fill in these values.

Model will be a two stream model : first stream gets the team embeddings and hand crafted features as input and second stream gets the 28 or 30 player embeddings as input, we late fuse the two streams and then predict the output vector of dim 9 +( 14/15 * 7 * 2)

Before we generate embeddings, we need to normalize the data in each of our columns, so that they are in the same range of values.

Besides the embeddings we need to add some hand crafted features:

1. Home/Away in the match stream
2. Form from league data / CL data ( Last 5-10 games) in match stream

The model output for the team should be:

1. Blocks_team_A
2. Blocks_team_B
3. Possession of one team
4. Passes_team_A
5. Passing_accuracy_team_A / Passes_completed_team_A
6. Passes_team_B
7. Passing_accuracy_team_A / Passes_completed_team_B
8. Corners_team_A
9. Corners_team_B

The model output for the individual player should be:

1. Goals. Return that 100 length vector.
2. Blocks
3. Fouls
4. Shots
5. Shots on target
6. Interceptions
7. Offsides

Therefore total number of output parameters = 9 + 14/15 * 7 * 2 (2 because players of both teams)

To generate the embeddings for match data we need:

1. Goals Scored
2. Attempts on target
3. Attempts off-target
4. Total blocks
5. Corners
6. Offsides
7. Possession
8. Passes
9. Passing accuracy
10. Total Distance covered
11. Recoveries
12. Total tackles
13. Clearances
14. Fouls
15. Cards

To generate Embeddings from player data we need:

1. Time played, i.e., finish time - start time.
2. Shots
3. Shots on target
4. Goals. We can pass a 100 length vector here with 1 in the index for the time at which the goal was scored.
5. Assists
6. Crosses
7. Interceptions
8. Fouls committed
9. Offsides