Enhancing Basketball Game Outcome Prediction through Fused Graph Convolutional Networks and Random Forest Algorithm

Area of Science:

• Sports Analytics
• Machine Learning
• Network Science

Background:

• Traditional machine learning models often overlook complex team dynamics in sports prediction.
• Vector-based models fail to capture the spatial structure and inter-team relationships within a league.

Purpose of the Study:

• To apply graph neural networks (GNNs) for predicting basketball game outcomes.
• To represent team interactions and league structure using graph-based data transformations.
• To improve upon existing prediction accuracy by incorporating relational data.

Main Methods:

• Structured basketball data (2012-2018 NBA season) was converted into undirected graphs.
• A graph convolutional network (GCN) was employed on the constructed team representation graph.
• Feature extraction using the random forest algorithm was integrated to enhance the GCN model.

Main Results:

• The initial GCN model achieved an average prediction success rate of 66.90%.
• The fused model, combining GCN with random forest features, improved prediction accuracy to 71.54%.
• The proposed GNN-based approach demonstrated superior performance compared to baseline models and previous studies.

Conclusions:

• Graph neural networks effectively model team interactions and spatial league structures for improved prediction.
• The integration of random forest feature extraction further boosts the predictive power of GNNs in sports analytics.
• This study offers a novel approach to basketball game outcome prediction, highlighting the potential of network science in sports research.