Sports match outcome prediction with graph representation learning

Predicting the winner of a sports match is a widely studied challenging problem of interest to many stakeholders (media, teams, fans, bookmakers). Various machine learning methods have been used in this field. However, the majority of past research has focused on learning about teams, not players. In other words, players' strength and lineup information, has not been considered in most previous works, despite its significant impact on match outcome prediction. In competitive team sports, the outcome can depend on complex interactions between opposing teams and interactions among players and teams. This thesis develops a novel approach to match outcome prediction that leverages graph representation learning to model team-team and team-player interactions. Both teams and players correspond to nodes in a Spatio-Temporal graph. Node embeddings capture how team/player characteristics jointly influence match outcomes. Empirical results on a dataset of nine different football leagues demonstrate the superior performance of our graph representation approach.