This competition was organized in the context of the "Introduction to Machine Learning" course (ELEN0062-1) at the University of Liège, Belgium. The goal was to let us apply the methods and principles exposed in the theoretical course in order to address a real problem: pass prediction during football matches.

The mission of our group, was to use (supervised) learning techniques to design a model able to predict the next player who would receive the ball via a pass based on the position of all players and the ball at a given time.

To achieve this, one provides us with a set of 8682 samples (passes) with the player positions, who are the sending and receiving players. A test set of 3000 unlabeled data was also provided for which it was ask to provide the next player who would receive the ball via a pass.

A more precise description of the data is available in the Data section.

Note that we could not use any information outside of those provided.

The provided data set has been (randomly) divided into a training set and a test set.

Each sample (corresponding to a pass) is a snapshot of the situation when the ball is passed, and contains the positions of the 22 players on the field (11 per team), the ID of the player (sender) with the ball, and the time (in ms, since the start of the half). The corresponding output is the ID of the player (receiver) that receives the passed ball.

Each player is represented by an ID from 1 to 22.

The pitch is 105 meters long (x-axis) and 68 meters wide (y-axis), and coordinates are given in cm. The coordinates of the center of the pitch are (0,0).

• input_train_set.csv - the input data of the training set
• output_train_set.csv - the output data of the training set
• input_test_set.csv - the input data of the test set

input_train_set.csv:

• Id - pass Id;
• time_start - number of milliseconds since the beginning of the concerned half-time period;
• sender - the ID of the player who has the ball - {1..22};
• x_<ID> - the x position of the player <ID> - [-5250 (cm); +5250 (cm)];
• y_<ID> - the y position of the player <ID> - [-3400 (cm); +3400 (cm)];

output_train_set.csv:

• receiver - the ID of the player who receive the ball via a pass [data to predict] - {1..22}

input_test_set.csv : see input_train_set.csv.

It has been decided to enrich our data set with some numerical features, mainly being distancesover the football field. Let’s review which features have been selected to solve the prediction problem. These features are computed by FeatureDerivation.

• Distance between the sender and the receiver: This feature was already implement by the toy_example.py provided file and has been kept. It simply consists of the Euclidean distance between the ball sender and the potential receiver (drawnamong all players on the field including the sender).
• Distances between the sender and his two closest team mates: these features are represented correspond to the distance between the sender of the pass and his two closest teammates.
• Distances between the receiver and his two closest opponents: These features has also been added and are the Euclidean distances between the receiver of the ball during the pass and his two closest opponents around him.
• Distances between the sender and his two closest opponents: These features has also been added and are the Euclidean distances between the sender of the ballduring the pass and his two closest opponents around him.
• Distance between the pass line and its closest sender’s opponent: An important feature for the team members was the smallest distance between a member of the sender’s opposed team and the pass line. Indeed, if some opponent are between the sender and the receiver, it seems logical that the pass is more risky and therefore that the sender will tend to choose another receiver.
• Distance between the sender/receiver and the field center: These features correspond to the Euclidean distances between the sender and the receiver of the ball during the potential pass and the center of the football field located in (0,0).
• Gain on the abscissa of the ball: This feature represents the horizontal movement of the ball towards the goal of the team opposed to the one of the sender. To compute this attribute, the side of the field occupied by each team has to be known. To determine this, the algorithm begins by determining the leftmost player on the field, considering that it is the goalkeeper of one of the 2 teams. If this player is on the same team than the ball sender, the goal in which he will have to score is on the right of the field, otherwise it is on the left. Note that here, an abstraction is made: a player will never be located behind the goalkeeper of the team on the left side of the field. It has been assumed that these situations are rare in football and thus that this is a right way to guess teams’ sides.
• Ordinate displacement of the ball: This new data has been added to the model by simply computing the absolute value of the difference of ordinates coordinates between the sender and the potential receiver of a pass.

• Predicate indicating if the sender is on the same team as the receiver of not: This feature was already implement by the toy_example.py provided file and has been kept. It simply consists of a boolean value being 1 (i.e. True) if the potential receiver of a pass is a team member of the sender, and 0 (i.e. False) otherwise.
• Predicate indicating if the sender’s team is performing an attack or not: In a football game, the pass performed can depend of whether the team possessing the ball performs an attack or not. It thus has been decided to implement a function able to determine if thesender is in an offensive or defensive phase of game. To do this, first the sender’s team field side is determined using the same method as described above. After that, the mean value of the x location of each player on the sender’s team is computed. Using these values, we can determine where most of the players in a team are located. If they are mostly in their camp, the action is considered as a defensive one, if they are not it’s an offensive one.
• Number of sender’s opponents around the receiver: This characteristic simply gives the number of opponents in a radius of 15,8 meters from the ball receiver. This distance has not been chosen by chance. Indeed it represents the mean length of a pass for the whole data set.
• Location zone of the sender and the receiver on the field: it has been decided to split the field into several zones. The following zones definition was decided
  • Zone 1 (center): x ∈ [−3200cm; 3200cm], y ∈ [−1750cm; 1750cm];
  • Zone 2 (left goal): x ∈ [−5250cm;−3200cm], y ∈ [−1750cm; 1750cm];
  • Zone 3 (right goal): x ∈ [3200cm; 5250cm], y ∈ [−1750cm; 1750cm];
  • Zone 4 (high flank): x ∈ [−5250cm; 5250cm], y ∈ [1750cm; 3400cm];
  • Zone 5 (low flank): x ∈ [−5250cm; 5250cm], y ∈ [−3400cm;−1750cm].

To compare different algorithms and their hyper-parameters, we performed the so-called test set method, which divides the learning data set into three parts, namely the learning set LS, the validation set VS and the test set TS. In the scope of the considered problem, this method consists of the following steps

1. We first randomly split the original 8682 passes data set into a 5210 passes LS and 1736 passes VS and TS, corresponding to 60 and 20 % of the whole data set respectively.
2. Considering a certain combination of features, we then train the set of candidate models on LS.
3. We evaluate each of these on VS by predicting the probability for each pair of players of VS to correspond to a pass, by deriving from those probabilities the predicted receiver’s Id foreach pass, and by comparing these results with the true output values corresponding to VS. Performing this evaluation allows us to select the best model based on its performance on VS.
4. We then train the selected model on LS+VS, and evaluate its performance on TS using the Scikit-Learn metric accuracy_score, which gives a performance evaluate of the considered model.
5. We finally train the selected model on LS+VS+TS to reach the best accuracy.

The model providing the best accuracy was a Random Forest for a depth = 15. The test set method has been achieved in order to predict the performance of the model, the predicted accuracy was 0.3698. Actually, this prediction was very accurate since the evaluation of the model on our test set gave us an accuracy of 0.3705.

The implementation was created on Python 3.8.5 using Sklearn libraries, make sure they are available on your machine before running the code. The models can be generated using

python [model]_test_set_method.py

where [model] is either random_forest for a random forest model, knn for a k-nearest neighbors model, adaboost for an AdaBoost model or MLP for a multilayer perceptron model. Note that all the model are evaluate with the test set method for differents hyper parameters values and generate the features by calling FeatureDerivation.py.

The competition follows the same guidelines as the previous projects, i.e., a written report and codes must be submitted before the deadline on the Montefiore Submission Platform. The competition will end a couple of days before the deadline in order to let you add your latest results in the report. You can find more information here.

This challenge is inspired from the Football Pass Prediction Challenge of the 5th Workshop on Machine Learning and Data Mining for Sports Analytics.