Improving Basketball Prediction Performance with Graph Attention Networks and K-means Clustering

This paper proposes a novel framework that incorporates Graph Attention Networks and a K-means-based feature selection method to predict the outcomes of basketball games. It creates a graph for a team, selects the representative features by different clustering configurations (K=3,4,5), and trains the GAT model to capture the inter-team complex relationships. The experimental results of the NBA dataset, from 2012 to 2018, provide the best accuracy of 0.7857 when K = 4 and the number of top features is chosen to be 5. It also performs much better compared to PCA and LASSO. Therefore, the clustering integrated with GAT elevates feature selection and uplifts predictive performance, which henceforth makes it strong in terms of sports analytics.