Football-Tactical-Pattern-Recognition

Introduction

This project is part of my Bachelor of Science (B.Sc) thesis, aimed at leveraging data mining techniques to analyze football tactics.

With the growing availability of sports data, particularly in football, automated tools are increasingly important for analyzing various game aspects. In this comparison-based study, I developed a system to detect offensive tactical patterns used by football teams. The system processes event-based football data, breaking it down into phases that are clustered based on spatial characteristics. The clustering is performed using the K-means algorithm combined with the Dynamic Time Warping (DTW) distance measure, which was found to be the most effective combination for clustering sequences. After clustering, the phases are analyzed using pattern mining techniques. The Vertical Maximal Sequential Pattern (VMSP) algorithm is employed to extract frequently occurring patterns, which are then ranked by relevance.

A visual dashboard is included in the project, allowing users to interactively explore the results. Users can select different algorithms for each component of the system, including clustering and pattern mining, enabling a flexible comparison of different methods and their outcomes.

FTPR Package

The FTPR (Football Tactical Pattern Recognition) package is a core component of this project, providing key functionality for preprocessing, clustering, and mining patterns from football event data. It is designed to be modular and flexible, enabling users to apply a variety of algorithms to analyze football tactics.

Key Features:

• Preprocessing: The package provides utilities for transforming raw football event data into structured formats, segmenting it into phases for deeper analysis.

• Clustering: Multiple clustering algorithms with different distance functions are supported, including:

  • K-means combined with the Dynamic Time Warping (DTW) distance measure, for grouping phases based on spatial and temporal characteristics.
  • Agglomerative Hierarchical Clustering, which allows for grouping phases in a tree-like structure, providing a different approach to discovering similarities between sequences.

• Pattern Mining: The package implements pattern mining techniques to detect frequent tactical sequences, with support for:

  • Vertical Maximal Sequential Pattern (VMSP), an efficient algorithm for finding maximal frequent sequential patterns allowing to find a compact representation of tactics.
  • CM-SPADE, a highly efficient algorithm for mining frequent sequences, offering another option for users to extract meaningful tactical patterns from football data.

• Visualization: FTPR includes visualization tools to represent clustering results and mined patterns, facilitating a clearer understanding of tactical insights.

Installation

Follow the steps below to install the necessary components for running this project:

1. Install FTPR Package

First, install the FTPR package by navigating to the ftpr directory and running the following command:

cd ftpr
pip3 install .

2. Install Other Dependencies

pip3 install -r requirements.txt

3. Download SPMF

Download the SPMF tool, a Java-based application for running sequence mining algorithms, from the official website. After downloading, place the spmf.jar file in the root directory of the project.

Usage

This project utilizes Statsbomb's ball-event data, specifically focusing on the Premier League 2015/16 dataset. Follow the following steps to run the project:

1. Download the data

To download a Statsbomb dataset you can either use download_data.py script or run notebook/download_data.ipynb notebook.

An example usage of download data script would be like this:

python3 download_data.py --data-dir data \
                         --events-dir events \
                         --competition-id 2 \
                         --season-id 27

--data-dir: The root directory of data.

--events-dir: The subdirectory of events within the root directory of data.

--competition-id: The competition id of the Statsbomb dataset.

--season-id: The season id of the Statsbomb dataset.

Alongwith the events, player.pkl file containing a dictionary of all players of all clubs will be stored in the root directory of data. Another file is matches.csv which contains the information of all matches in the dataset.

2. Extract the Phases

To extract the team phases you can use extract_teams_phases.py script. You can use this script as follows:

python3 extract_teams_phases.py --matches data/matches.csv \
                                --events data/events \
                                --output-dir data/team_phases

--matches: The path to matches.csv file.

--events: The path to subdirectory of events within the root directory of data.

--output-dir: The directory where the phases will be saved in.

3. Run the Dashboards

There are 2 dashboards available in this project. The first one dashboard/main.py is the main dashboard where you can run clustering and sequence mining algorithms. The second one dashboard/phases.py.

To run the main dashboard simply run:

cd dashboard
python3 -m streamlit run main.py -- --rows 4 --cols 3 \
                                 --players-path ../data/players.pkl \
                                 --matches-path ../data/matches.csv \
                                 --spmf-dir ..

--rows: Number of rows of clusters shown in each page.

--cols: Number of columns of clusters shown in each page.

--players-path: The path to players.pkl file.

--matches-path: The path to matches.csv file.

--spmf-dir: The path to spmf.jar file.

To run the phases dashboard simply run:

cd dashboard
python3 -m streamlit run phases.py -- --phases-dir ../data/team_phases \
                                 --matches-path ../data/matches.csv

--phases-dir: The directory where the phases are saved in.

--matches-path: The path to matches.csv file