An analysis of player-formation graphs for predicting football passes

Stropnik, Vid (author), Šubelj, Lovro (mentor)

In modern association football, data and the knowledge derived from it play a crucial role in forming tactical plans and analyzing games. In this Master's thesis, we explore the capability of modeling a passer's decision-making when selecting a target during a football match, using graphs and graph neural networks. We present a methodology for constructing a dataset of graphs from open data, where nodes represent players on the field and edges represent interactions between them. We investigate three different graph configurations and analyze the utility of advanced features, such as player movement trajectories and pitch control models. Our final dataset contains 80,332 graph representations from 166 matches. We evaluate the usefulness of the dataset on three tasks: regression of the coordinates of a successful pass, classification of the role of the pass recipient, and prediction of the target zone of the pass, based on the positional play pitch division. Our models, based on the Graph Transformer architecture, achieve the best results in all tasks. The results demonstrate that our approach effectively models the passer's decision-making and that graph models enriched with appropriate features can significantly contribute to advanced football analytics.

• Language: English
• Keywords: association football, soccer, graph neural network, graph dataset
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-164917
• Publication date in RUL: 15.11.2024

Secondary language

• Language: Slovenian
• Title: Analiza grafov postavitev igralcev za napovedovanje podaj na nogometnih tekmah

Abstract

V modernem nogometu imajo podatki in iz njih pridobljeno znanje ključno vlogo pri oblikovanju taktičnih načrtov in analizi iger. V tej magistrski nalogi raziskujemo zmožnost modeliranja podajalčevega odločanja pri izbiri tarče med nogometno tekmo z uporabo grafov in grafovskih nevronskih mrež. Opišemo metodologijo za konstruiranje podatkovne zbirke grafov iz odprtih podatkov, v kateri vozlišča predstavljajo igralce na igrišču, povezave pa interakcije med njimi. V delu med seboj primerjamo tri različne grafovske konfiguracije in analiziramo uporabnost naprednih značilk, kot so vektorji igralčevih premikov ter domenski modeli nadzora igrišča. Naša končna podatkovna zbirka vsebuje 80,332 grafovskih ponazoritev, vzorčenih na 166 tekmah. Uporabnost podatkovne zbirke evalviramo na treh nalogah: regresiji koordinat uspešne podaje, klasifikaciji vloge prejemnika podaje in napovedovanju ciljnega območja glede na specifično delitev igrišča. Naši modeli, osnovani na arhitekturi grafovskega transformerja, dosegajo najboljše rezultate pri vseh nalogah. Rezultati kažejo, da naš pristop učinkovito modelira podajalčevo odločanje in da lahko grafovski modeli, obogateni z ustreznimi značilkami, pomembno prispevajo k napredni nogometni analitiki.

• Keywords: nogomet, grafovska nevronska mreža, grafovska podatkovna zbirka