Footballer Player Recommendation Model Using Graph Convolutional Networks

In modern football, smart recruitment is essential in building successful teams. Traditional scouting methods, though effective in identifying player mentalities, often fall short due to the team’s budget restrictions and the need to fit the head coach’s strategy. Teams frequently lose top targets to wealthier clubs, sign players who do not fit with the coach’s strategy and overlook affordable and stylistically similar players. To address this, we propose a novel player recommendation system, which first clusters players and then leverages the Graph Convolutional Networks (GCNs) ability to generate rich node embeddings from a player-similarity graph constructed via cosine distance. Using data from FBref, we refine the scope to only central midfielders and on the Graph Convolutional Network where the target variables are the labels from K-Means clustering of which we obtained 49% accuracy with all neighbours but a perfect 100% with shared-label aggregation introduced in this paper. It was observed that there were significant visual differences (obtained via t-SNE) in the clusters formed. Graph Convolutional Networks are primarily used for node classification on academic citation datasets, and the proposed method in this research outperformed on the Cora (88%) and Pubmed (84%) Datasets, respectively.