In basketball, knowing the odds of success of a shot is critical. Many factors affect shot success, including the location of the shot, the shot style (e.g., jump shot, finger roll, dunk), and of course the player`s skill. Crucially, the list of highprobability shots is different for each player. We predict the success of shot made by NBA basketball players in the 2015-2016 season, and show that using traditional methods such as logistic regression or support vector machine regression is problematic because many style-location-player combinations do not occur in the training data, i.e., the data are highly sparse. Hence, we propose a shot recommender system based on factorization machines. Factorization machines have been used successfully in recommendation problems because they handle sparse data, scale well to very large datasets, and provide latent factors that capture underlying rater (player) preferences and item (shot) features. For the NBA player data, a 25-factor model predicts logodds of shot success with high accuracy. It also identifies both highly recommended shots and shots to be avoided, including shots that are not represented in the training data.
© Copyright 2016 Proceedings of the KDD-16 Workshop on Large-Scale Sports Analytics. Veröffentlicht von Eigenverlag. Alle Rechte vorbehalten. / Übersetzt mit https://www.DeepL.com/Translator