Comparing Traditional and Streaming Machine Learning Methods for Soccer Pass Detection - Thesis Abstract

Aug 2, 2026

Sports analytics increasingly relies on continuous data streams from wearable tracking devices, yet traditional Machine Learning struggles with incremental learning as data evolves. Streaming Machine Learning addresses this by enabling continuous learning and adaptation to incoming data. This thesis compares both approaches for soccer pass detection, classifying pass vs non-pass actions using leg-movement features extracted from sensors on players’ shoes. Balanced, progressively imbalanced, and resampled datasets (SMOTE, C-SMOTE) are evaluated with standard metrics, while t-tests and Nemenyi tests assess significance. Results show streaming frameworks can match or exceed traditional models, with ensemble methods delivering strong performance and streaming methods outperforming on larger rebalanced datasets.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages110

Reading Time4 h 37 min

AccessFree to download

Abstract

Abstract in Italiano

1 Introduction

1.1 Thesis Contributions

1.2 Thesis Organization

2 State of the Art

2.1 Sports Analytics

2.1.1 DEBS 2013 Grand Challenge

2.2 Machine Learning

2.2.1 Decision Tree

2.2.2 Bagging

2.2.3 Random Forest

2.3 Streaming Machine Learning

2.3.1 Hoeffding Tree

Comparing Traditional and Streaming Machine Learning Methods for Soccer Pass Detection Tesi di Laurea Magistrale in Computer Science and Engineering Ingegneria Informatica Author: Stefania Menconi Student ID: 967553 Advisor: Prof. Emanuele Della Valle Co-advisors: Alessio Bernardo, Giacomo Ziffer Academic Year: 2022-2023 i Abstract Sports analytics has grown significantly through continuous data streams from wearable tracking devices. However, traditional Machine Learning methods struggle with incremental learning from evolving data. Streaming Machine Learning enables models to continuously learn and adapt from incoming data. This thesis aims to compare the performance of traditional and Streaming Machine Learning approaches in sports analytics by identifying passes during a soccer match. We evaluate the ability of algorithms from both fields to differentiate between pass and non-pass actions using only leg movement features extracted from sensors placed on players’ shoes. Initial balanced and progressively imbalanced datasets are generated alongside larger rebalanced datasets produced statically via SMOTE and dynamically via C-SMOTE. Traditional and streaming algorithms such as decision tree, bagging, random forest, Hoeffding tree, leveraging bagging, adaptive random forest, and streaming random patches are tested on these datasets using a variety of standard performance metrics. Statistical analyses like the t-test and Nemenyi test determine significant differences between algorithms’ performance. Results demonstrate that Streaming Machine Learning frameworks can perform on par with or surpass traditional Machine Learning in various contexts. While traditional techniques initially hold advantages, streaming algorithms consistently outperform traditional counterparts over larger rebalanced datasets. Ensemble methods prove particularly robust across paradigms, and additionally, leveraging bagging and streaming random patches emerge as top performers, though adaptive random forest maintains competitiveness. This work provides compelling evidence that Streaming Machine Learning achieves comparable or better performance than traditional Machine Learning. Leveraging streaming techniques’ adaptability extends possibilities for online sports analytics. Future works expand analysis through varied matches modelling player tendencies and evaluate algorithms’ long-term adaptability on edge devices. Overall, this research underscores Streaming Machine Learning’s growing relevance for analyzing continuous data streams in dynamic domains like sports. Keywords: Machine Learning, Streaming Machine Learning, Sports Analytics, Imbalanced Classification, Resampling, Statistical Tests Abstract in Italiano L’analisi sportiva ha visto un notevole sviluppo grazie ai flussi continui di dati provenientidai dispositivi indossabili di tracciamento. Tuttavia, i metodi tradizionali di Machine Learning faticano nell’apprendimento incrementale da dati in evoluzione. Streaming Machine Learning consente ai modelli di apprendere in modo continuo e adattarsi ai dati in arrivo. Questa tesi si propone di confrontare le prestazioni dei metodi tradizionali e di Streaming Machine Learning nell’ambito dell’analisi sportiva, attraverso l’identificazione dei passaggi di una partita di calcio. Valutiamo la capacità degli algoritmi di entrambii campi di differenziare tra azioni di passaggio e di non passaggio, utilizzando solo le caratteristiche del movimento delle gambe estratte da sensori posizionati sulle scarpe dei giocatori. Vengono generati dataset iniziali bilanciati e progressivamente sbilanciati, insieme a dataset più ampi ribilanciati staticamente tramite SMOTE e dinamicamentetramite C-SMOTE. Diversi algoritmi tradizionali e in streaming come decision tree, bagging, random forest, Hoeffding tree, leveraging bagging, adaptive random forest e streaming random patches sono testati su questi dataset utilizzando una varietà di metriche d

FAQs

What problem does the thesis address in soccer pass detection?

It addresses how to distinguish pass actions from non-pass actions in soccer using features extracted from wearable leg-movement sensors on players’ shoes.

How does Streaming Machine Learning differ from traditional Machine Learning in this work?

Streaming Machine Learning continuously updates and adapts as new incoming data arrives, while traditional methods are evaluated in settings that do not inherently support incremental adaptation.

How are imbalanced datasets handled and evaluated?

The study uses balanced and progressively imbalanced datasets, plus larger rebalanced datasets created statically via SMOTE and dynamically via C-SMOTE. Algorithms are then compared using standard performance metrics and significance tests including the t-test and Nemenyi test.

Aug 2, 2026

Upload by Evangeline