Intelligent Prediction of the Sport Game Outcome Using a Hybrid Machine Learning Model

The National Collegiate Athletic Association (NCAA) serves as the platform for showcasing the skills of talented basketball players from various colleges. With the historical set provided by NCAA this study proposes a hybrid model which is combining the gradient boosting decision tree (GBDT), Tabnet and support vector machine (SVM) for 2023 NCAA basketball game outcome. For each possible matchup between two top college teams, the model can predict the probability of the win rate and the winner team. The fusion model combines the strengths of tree-based model, linear models like SVM and Tabnet to enhance prediction performance, robustness, and interpretability. The data exploration and preparation part shows the important features like the win Ratio of different teams and the feature engineering for the further model training. The experiment part shows the data distribution and feature engineering and performance for each model. The hybrid model beats the separated model with a better brier score of 0.176, which shows the superiority of the hybrid model. Keywords: basketball game; GBDT; hybrid model; NCAA; SVM; Tabnet