Code
Web
The NBA is one of the most widely viewed sports on the planet, with 30 teams competing to qualify for the playoffs and eventually win the season title. Predicting playoffs qualification and analyzing performances of the team as a whole are imperative to identify areas of improvement for the teams. However, a detailed study reveals that these areas either lack formal work or involve older traditional ML techniques.
In this work, we introduce a novel architecture PlayoffsNet to predict playoffs qualification of NBA teams. Key points in this study are:
- Detailed study of the data, analyzing changes and evolution of the sport and its regulations over the years, and the distinction between qualifying and non-qualifying teams.
- Elaborate and exhaustive analysis on the choice of activation function for the architecture – ReLU, Leaky ReLU, TanH, Swish, Mish.
- Explaining the contributing factors towards positive and negative predictions using SHapley Additive exPlanations (SHAP).
The accuracy, F1-Score, and loss of the model are aggregated over N=25 iterations, and the stable mean of peaks for each metric is calculated using:-
Our model achieves an improvement of about 11% over the existing state-of-the-art models for NBA playoffs prediction achieving a peak accuracy of 86.4%. It also shows the top factors influencing the outcome of the model, and enables obtaining actionable insights on the identifying potentially weak areas of performance for the team as a whole. Users can interact with our model online and pass in their favorite team and season to understand whether that particular roster would qualify for the playoffs.
Kurban Intelligence Lab 