The THESIS_NOTEBOOK contains all the codes and the functions used in order to define the technical features of a football team and to compute the machine learning classifiers used to classify a team in a game as male or female.

How to use it:

• 1 STEP: download the soccer-logs data of interest, collected from https://www.nature.com/articles/s41597-019-0247-7, into a folder (which we call data, for example);

• 2 STEP: use data_import.sh file into the Command prompt to create the MongoDB dati database, where I storage the usefull data;

• 3 STEP: obtain the PlayeRank functions using git clone https://github.com/mesosbrodleto/playerank.git; NOTE before computing the players' ratings in SECTION 6.3 of the THESIS_NOTEBOOK, you first need to define the feature_weights.json and the role_matrix.json files. These files are obtained by launching the codes at the beginning of SECTION 6.3 of the THESIS_NOTEBOOK and they use the information of events, players and matches into the data folder.

After you do these preliminar phases, you can use the THESIS_NOTEBOOK!

This work uses data related to the Fifa Men's World Cup 2018, with 101,759 events from the 64 games played and 736 players, and to the Fifa Women’s World Cup 2019, with 71,636 events on the 44 matches available and 546 players.

Again, all the data are storage in a MongoDB database which I called dati and they are presented in json format, i.e., characterized by collections and documents. A collection is characterized by documents that describe: a World Cup game, a team, a player or a soccer event, i.e., a certain action, such as a pass, a shot, a foul, a save attempt, and so on, made by a team's player in a match.

NOTE the dati database it's been created by me and contains all the usefull collections; they were imported through the 'mongoimport' command in the Command prompt. See data_import.sh file to create the dati database and import all the usefull json files/collections.

The dati database contains the following collections:

• teams (male football teams)
• teamsW (female football teams)
• players (male players)
• playersW (female players)
• eventsWC (events in the Russia 2018 World Cup championship)
• eventsWCW (events in the France 2019 World Cup championship)
• matchesWC (matches in the Russia 2018 World Cup championship)
• matchesWCW (matches in the France 2019 World Cup championship)

A document describing a match, for example, presents the identities of the two challenging teams through the variable teamsData, which contains the variable formation (list of sub-documents) with the description of the players on the bench, on the lineup and information on any substitutions. For each player the following features are described: playerId, assists (dummy), goals, ownGoals (dummy), redCards (dummy) e yellowCards (dummy). Finally, the document presents generic data on the match and its outcome (winner, date, referees, etc...).

Of particular importance are the documents describing an event, such as:

{'_id': ObjectId('5d8a3d05f89835179f701044'), 'eventId': 8, 'subEventName': 'High pass', 'tags': [{'id': 1801}], 'playerId': 139393, 'positions': [{'y': 53, 'x': 35}, {'y': 19, 'x': 75}], 'matchId': 2057954, 'eventName': 'Pass', 'teamId': 16521, 'matchPeriod': '1H', 'eventSec': 4.487814, 'subEventId': 83, 'id': 258612106}

Which describes a High pass made by the Arab player Abdullah Ibrahim Otayf during the game "Russia - Saudi Arabia", after about 4 seconds from the kick off.

We decide to use Python as MongoDB driver, via the pymongo package. All the collections into dati are renamed as teamsM, teamsF, playersM, playersF, eventsM, eventsF, matchesM and matchesF, respectively.

All the function we use to create the different technical features of a football team are in the .py files: (i) match_intensity.py (ii) H_indicator.py (iii) flow_centrality.py .

---

In Section 6.3 we quantify the performance ratings of each player in a certain game, and then we aggregate them via mean and standard deviation, for each national team. The ratings are computed using the PlayeRank framework, taken from:

• Playerank -- Pappalardo, L., Cintia, P., Ferragina, P., Massucco, E., Pedreschi, D., & Giannotti, F. (2019). PlayeRank: data-driven performance evaluation and player ranking in soccer via a machine learning approach. ACM Transactions on Intelligent Systems and Technology (TIST), 10(5), 59.

• Public dataset -- Luca Pappalardo, Paolo Cintia, Alessio Rossi, Emanuele Massucco, Paolo Ferragina, Dino Pedreschi & Fosca Giannotti (2019). Nature Scientic Data.

To implement it, you need to use json files and to compute the features' weights and the role_matrix as descripted in the articles. They also need to be in json format. The weights are estimated via SVM and they indicate the importance of every event(in the winning goal) that occurred in a football match. While, they use a k-means algorithm to detect the player roles, based on the average position of them with respect to their mates on the field (center of performance).

To compute team players' ratings of a certain soccer league or tournament, you need the json files describing matches, events and players of that competition; for example, we use matchesWcup_Women, eventsWcup_Women and playersF to compute the performance ratings of the female players who have played in the France 2019 World Cup.

Finally, we aggregate ratings via mean and standard deviation, for each national team and for every game. Ratings' standard deviation of a team quantifies how important the indivual performance of a player affects the team collective performance; for example in the match Portugal vs Spain (3-3, id=2057960), the ratings' standard deviation of Portugal is quite high, Rstd=0.22, due to the great individual performance of C. Ronaldo that scored 3 goals.

---

In Chapter II, we implement the machine learning classifiers to classify a football team as male (class 0) or female (class 1), in a certain game. We use the models: Logistic classifier, Decision Tree, Random Forest and Adaboost classifier (and also a Baseline which assign always the most frequent class, i.e., male team). We carry out two different experiments:

• in the first, we select the tuning parameters of each method through the 'GridSearchCV' function, taken from the sklearn library, on the 20% of the total observations available, randomly drawn. There are chosen the parameters that maximaze the 5-fold CV accuracy. After, we validate the models via 10-fold CV, computing the accuracy and F1-score, on the remaining 80%.

NOTE Due to a random component present in each classifier, the choice of the best parameters may sometimes slightly differ from the choice of parameters presented in the notebook.

• in the second experiment, we split again the training set into another training set (70%), which we call fitting set, and into a test set (30%). This split is repeated N times (twenty times in our case) based on different random state values, and we fit the Decision Tree with the best parameters on the fitting set each time. Every time the Decision Tree is fitted, we take the feature importance measure for each variable, and we measure the Normalized Root-Mean-Square Error and the Kendall's τ coefficient between each pair of trees, to stabilize on which features the Decision Tree focuses on the most to classify a team in a game as male (class 0) or female (class 1).

NOTE If you want to focus on just a subset of variables, you can put your list of variables of interest instead of 'dref', and compute these two indices to evaluate the grade of concordance between every pair of trees, about the importance of that specific variables in the classification task. For example, we use the features: ratings' standard deviation, average pass velocity, ball possession recovery time and percentage of accurate passes.