A Hybrid Approach for Interpretable Game Performance Prediction in Basketball

Sports Data Analytics is an emerging field involving analytical tools to help coaches design practical strategies for athletic training and winning games. In this paper, we developed Decision Tree (DT) based models for analyzing the sleep and recovery data, training statistics and cognitive state of athletes in a Women's Division-I basketball team for game score prediction. We develop a hybrid approach that employs classification/regression trees and random forests to predict the weighted game score in conjunction with factor analysis. These factor weights are explained using a consensus score derived from a small collection of decision trees constructed during the prediction. These decision trees on further inputs from coaches can lead to a robust mechanism for interpreting model predictions. Our athlete�s data consisted of 2800 records, with 37 attributes obtained from 16 athletes over 25 weeks from strength coaches, sleep and recovery information from wrist-worn devices, and perception questionnaires. Our hybrid approach first performs factor analysis to compactly characterize the athlete's data in terms of 7 latent factors. It then uses these factors to build a collection of decision trees for predicting and interpreting the game score. The predictions from our approach have an MSE of 0.013 and an R
²
of 0.971, which is better than classical multilinear regression (MSE 0.053, R
²
0.689) and DT (MSE 0.036, R
²
0.798) based approaches.