Blog of Random Thoughts and Pictures

Player Ranking Framework

April 11th, 2021

As mentioned in Part I of this multipart post, Luca Pappalardo prepared a video, for the Friends of Tracking channel in 2020, to talk about some elements of a paper related to an open Wyscout data set, and advanced statistics related to passing networks, flow centrality and player ranking.

For this post (Part IV) I’m going to cover my take on the PlayeRank framework created by this team of researchers.

I’ve forked the “mapping-match-events-in-Python” repo into my mmoffoot area and created a new branch called ‘englanddata’ to cover the data set of English Premier League information for the 2017-18 season.

An exhaustive description of the PlayeRank framework is available in this paper Pappalardo, Luca, Cintia, Paolo, Ferragina, Paolo, Massucco, Emanuele, Pedreschi, Dino & Giannotti, Fosca (2019) PlayeRank: Data-driven Performance Evaluation and Player Ranking in Soccer via a Machine Learning Approach. ACM Transactions on Intelligent Systems and Technologies 10(5).

This Notebook builds player rankings from match events, the following steps are required:

  • compute feature weights (learning)
  • compute roles (learning)
  • compute performance scores (rating)
  • aggregate performance scores (ranking)

It doesn’t take long to run through the [In] steps of the Notebook and for the English data you end up with Figure 1 as seen below.

Figure 1: Player Ranking English Premier League 2017-2018

The visual output from the Notebook is interactive which is great as you can hover over the points to catch the name. For example in the striker role H. Kane is the outlier (at the top), S. Augero second. There’s even a drop down menu to do a comparison.

Figure 2: Player Ranking Comparison for H. Kane and S. Augero English Premier League 2017-2018

The positions are an interesting element to this ranking systems which is based on a role matrix.

Team Attacking left to right, position 0 is a Striker

And the top players from English Premier League 2017-2018 for each role position

  • Position 0. = H. Kane
  • Position 1. = L. Milivojevic
  • Position 2. = N. Monreal
  • Position 3. = D. Janmaat
  • Position 4. = S. Mane
  • Position 5. = M. Salah
  • Position 6. = N. Otamendi
  • Position 7. = J. Stephens

If I look at the PFA Premier League Team of the Year for 2017-18, Otamendi, Kane and Salah were named in it and also appear here in PlayeRank, but none of the rest. I wonder how D. Silva and K. De Bruyne both of whom are in the PFA team, missed out in this PlayeRank framework.

Overall over 4 posts I can say this is a great Jupyter Notebook, firstly to really learn about Jupyter Notebooks, and secondly to be able to see the structure and how to use WyScout data. It is so important given this data set is used by so many tops clubs for the scouting, analyses and recruitment of players.

I got caught a little on the passing networks, and the flow centrality but certainly a thread of more investigation on a measure of cohesiveness within the team, would be a nice continuation of this topic.

The player ranking and the full explanation of the PlayeRank Framework was fantastic and a joy to read and interact with.

Passing networks and Flow centrality

April 5th, 2021

As mentioned in Part I of this multipart post, Luca Pappalardo prepared a video, for the Friends of Tracking channel in 2020, to talk about some elements of a paper related to an open Wyscout data set, and advanced statistics related to passing networks, flow centrality and player ranking.

For this post (Part III) I’m going to cover my take on Passing networks and Flow centrality.

I’ve forked the “mapping-match-events-in-Python” repo into my mmoffoot area and created a new branch called ‘englanddata’ to cover the data set of English Premier League information for the 2017-18 season.

Passing networks

This Notebook creates a player passing network for any of the matches covered in the data set. The passing network is a weighted network where nodes are players and weighted edges represent movements of the ball between players. The size of an edge is proportional to the number of passes between the players.

Some finer details on are covered in a research paper Cintia et al., The harsh rule of the goals: data-driven performance indicators for football teams, In Proceedings of the 2015 IEEE International Conference on Data Science and Advanced Analytics (DSAA’2015), 2015.

Now I wanted to continue with a passing network from the Tottenham Hotspur – Leicester City match (2018), given this was the match showing the events to this point, however it just turned out as a blank image for both teams via the Notebook, I tried another match and only got one team and finally the Arsenal – Burnley match turned out what was being looked for, so I moved over to this match.

Arsenal Passing Network, Arsenal – Burnley, May 13, 2018
Burnley Passing Network, Arsenal – Burnley May 13, 2018

The produced images need some time for study, even a quick look leaves me wondering what I can take from them. I also tried to gain some insight from a related paper from the authors P. Cintia, S. Rinzivillo, and L. Pappalardo, “A network-based approach to evaluate the performance of football teams,” in Proceedings of the Machine Learning and Data Mining for Sports Analytics workshop, ECML/PKDD 2015, were it mentions again that nodes are players, directed edges represent passes between players and the size of an edge is proportional to the number of passes between the players. Node 0 indicates the opponent’s goal, and edges ending in 0 node represent goal attempts. However there are no Node 0 in these passing networks.

The related video at minute 15:00 has a section that describes the passing networks too, but unfortunately there’s still not enough in there to give a hint as to what the take away is.

Flow centrality

Next up is flow centrality, which is a feature that can be computed on the passing network and in this Notebook is described as a way to capture the fraction of times that a player intervenes in those paths that result in a shot. They take into account defensive efficiency by letting each player start a number of paths proportional to the number of balls that he recovers during the match.

This concept is only lightly explained in the paper Duch et al., Quantifying the Performance of Individual Players in a Team Activity, PLoS ONE 5(6): e10937 as referenced in the Notebook and I must admit I still didn’t get it, but then a read of a source paper by Freeman LC, A set of measures of centrality based upon betweenness, Sociometry 40: 35–41, 1977 clears it all up “when a particular person in a group is strategically located on the shortest communication path connecting pairs of others, that person is in a central position”, and that same paper goes on to show measures that define centrality in terms of the degree to which a point falls on the shortest path between others and therefore has a potential for control of communication.

Flow Centrality Burnley 2018, Arsenal – Burnley May 13, 2018

Dare I say Westwood (Burnley) is the betweenness man in control for Burnley.

Might be a follow up idea to classify based on how they measure cohesiveness within the team.

Advanced football visualisations duels on the pitch Italy compared to England

March 27th, 2021

As mentioned in Part I of this multipart post, Luca Pappalardo prepared a video, for the Friends of Tracking channel in 2020, to talk about some elements of a paper related to an open Wyscout data set, and advanced statistics related to passing networks, flow centrality and player ranking.

For this post (Part II) I’m going to cover my take on the match evolution and spatial stats.

I’ve forked the “mapping-match-events-in-Python” repo into my mmoffoot area and created a new branch called ‘englanddata’ to cover the data set of English Premier League information for the 2017-18 season.

Spatial distribution of events

There are tons of events collated in the WyScout API Events from duels to fouls to interruptions, as explained in the WyScout API document. For passes there are 6 types:

  • Pass = Hand pass,
  • Pass = Head pass,
  • Pass = High pass,
  • Pass = Launch,
  • Pass = Simple pass,
  • Pass = Smart pass,

In this Notebook there’s an interesting set of images created which show the distribution of positions per event type. These kernel density plots show the distribution of the events’ positions during the match with the darker the green representing the higher number of events in a specific zone of the field.

Figure 1: Move the slider to compare Italian Serie A Duels and English Premier League Duels 2017-18

The first image is of duels, and in the WyScout world “Duel” has a specific meaning,

A challenge between two players to gain control of the ball, progress with the ball or change its direction.

With a number of subtypes to consider too: Defensive duel, Offensive duel, Aerial duel, Loose ball duel, and Sliding tackle

For the moment I’m not going into the whys and wherefores of these subtypes, but it’s really interesting to review and compare the images and to see the difference of where the Italian league and the English league host their duels. Dare I say right-backs, left-backs and right wingers, left wingers should look closer if they are moving between the leagues.

A big summer move from England to Italy was Emre Can from Liverpool to Juventus, with Stephan Lichtsteiner coming in from Juventus to Arsenal maybe a view of these plots before the new 2018-19 season got underway might have been handy.

Of note there’s a 10,000 event sample size in here by default, so for the Italian & English league this represents about 6 matches worth of events, and so a larger sample size would be nice to see and compare against. Would also be nice to identify specific players (RB,LB and RW, LW) that were strong in those main duel locations, however that will have to be for another day.

Here are where the fouls happen.

Figure 2: Move slider to compare Italian Serie A Fouls and English Premier League Fouls 2017-18

And the shots.

Figure 3: Move the slide to compare Italian Serie A shots and English Premier League shots 2017-18

Intra-match evolution of the events

Goals are the main stay of football and so when looking at the English and Italian leagues (season 2017-18), its pleasing to see the difference between the leagues, especially the 1st half goals.

Yellow cards and red cards are covered in the data set too, and displayed in the Jupyter Notebook but I’ll be honest and say I didn’t take too much time to analyse the results here, because I was fascinated by the Duel plots.

Advanced football visualisations and data analysis of match events

March 22nd, 2021

Luca Pappalardo an author of the paper (PCR2019) Pappalardo, L., Cintia, P., Rossi, A. et al. A public data set of spatio-temporal match events in soccer competitions. Nature Scientific Data 6, 236 (2019) prepared a video, for the Friends of Tracking channel in 2020, to talk about some elements of this paper and the related Wyscout data set, which was used for the paper.

In this video Luca covers:

  • The Wyscout data set, how it is collected, from players to events.
  • Basic statistics on events and distributions.
  • Plotting events on the field, match evolution and spatial stats.
  • Advanced statistics: passing networks, flow centrality and playerRank

For this blog post I’m going to cover my take on the player events in the Wyscout data set and the display of some basic statistics on events and distributions.

The origin of the code is available on Github under the project “mapping-match-events-in-Python” and worked through in this video.

Set up

Then I created a new branch called ‘englanddata’ in this area to cover the changes I made.

The example code base uses the Italian league data, but the branch name might be a give away, seeing as the data set has English Premier League information for the 2017-18 season I wanted to run the code base against that data set, and so I took a copy of the original Jupyter Notebook and ran it against the English data as data_england_exploration.ipynb.

The full list of data available includes:

  • Italian first division 2017-18
  • English first division 2017-18
  • Spanish first division 2017-18
  • French first division 2017-18
  • German first division 2017-18
  • European Championship 2017-18
  • World Cup 2018

All the matches, events, players, and competition data sets are hosted in a figshare repository with all the data stored in a JSON format.

The way the data is collected is explained in the paper, with a nice visual representation in the Notebook so I won’t ruin that insight and will let you read it in there.

I should say a quick word on Jupyter Notebooks, its an interactive way of developing and presenting data science projects, and I can really see that it’s an easy way to follow the code base for this project. It’s easy enough to install Jupyter Notebook on a machine too and well worth the install.

Plotting events on the field

There are a number of nice overviews of the structure of data given in the early part of the Notebook, but it’s more interesting when it comes to the static plots.

Figure 1: All Events Tottenham Hotspur 5 – 4 Leicester City, May 13, 2018.

Although of course too much detail can overwhelm and so the interactive plots in this Notebook are much better mechanism to share this information, as in you just have to hover the mouse over the event and its details come to the fore.

Figure 2: Pass Events Tottenham Hotspur – Leicester City, May 13, 2018.
 
Figure 3: Foul Events Tottenham Hotspur – Leicester City, May 13, 2018.
 
Figure 4: Fouls by a specific player Tottenham Hotspur – Leicester City, May 13, 2018.

This is a great Jupyter Notebook, firstly to really learn about Jupyter Notebooks, and then of course to be able to see the structure and how to use WyScout data. It is so important given this data set is used by so many tops clubs for the scouting, analyses and recruitment of players.

There’s more to come, as I plan to complete the match evolution, spatial stats in part II of this blog post and finally cover the advanced statistics: passing networks, flow centrality and playerRank in a part III of this blog post.