Using Deep Learning to Understand Patterns of Player Movement in the NBA

Research Paper will be posted in the coming weeks. Check back soon!
Download the
Full Paper Here

Akhil Nistala
John Guttag


Abstract: In 2011, SportVU fundamentally changed the way that basketball can be analyzed. STATS SportVU utilized a six-camera system installed in basketball arenas to track the real-time positions of players, at 25 frames per second. In this paper, we demonstrate how we can apply deep learning techniques to this data to produce a queryable database of basketball possessions.

We trained an unsupervised machine learning pipeline that generates a representation, called a trajectory embedding, of how individual players move on offense. The representation is a 32-dimensional vector of floating-point numbers that captures the semantics of a single player’s movement, such as locations of the endpoints, screen actions, court coverage, and other spatial features. We generated nearly 3 million trajectory-embeddings from three seasons of data (2013-2014, 2014-2015, 2015-2016).

We found that the Euclidean distance between trajectory-embeddings is an excellent indicator of the visual similarity of the movements they encode. For example, two different movements of a post-up in the right block will have nearby embeddings; a post-up in the right block and a screen action above the left wing will have distant embeddings. This result led to the Similar Possessions Finder, a queryable database of basketball possessions.

The Similar Possessions Finder can be used to quickly answer queries such as “How much more frequently did Andre Drummond establish position on the right block than on the left block during the 2015-2016 regular season?” and “Find all possessions from the 2014 playoffs in which Chris Paul ran a screen action in the high post that ended with DeAndre Jordan scoring.”