Using deep learning to understand patterns of player movement in basketball

Abstract

National Basketball Association (NBA) coaches spend a great deal of time analyzing the effectiveness of various strategies. Typically, this entails countless hours pouring over videos of games, and trying to derive generalizable conclusions from hundreds of thousands of examples. In this thesis, we present a methodology for quantitatively approaching this task. We start from player tracking data that records the position on the court of each player 25 times per second. We use an unsupervised machine learning pipeline to learn a low-dimensional encoding for each player's movement, over one possession on offense. Each encoding captures the semantics of a single player's movement, such as locations of the endpoints, screen actions, court coverage, and other spatial features. We generate 3 million such trajectory-embeddings from 3 seasons of data. These can be clustered to reveal trends in player movement between sets of games. Our framework can be used to answer such questions as "How did Klay Thompson's movements change between wins and losses during the 2016 NBA Finals?" (18% of his trajectories in wins were movements between the sidelines and corners, compared to 3.5% in losses) and "How much more frequently did Andre Drummond establish position on the right block than the left block during the 2015-2016 regular season?" (Almost 40% of his trajectories from 2015-2016 were right of the basket, compared to less than 15% to the left)..