A weighted network clustering approach in the NBA

Abstract

Evaluating players’ performance for decision-makers in the sports industry is crucial in order to make the right decisions to form and invest in a successful team. One way of assessing players’ performance is to group players into specific “types”, where each type represents a level of performance of its players within. In this paper, we develop a novel clustering approach in order to cluster types of players in the NBA. The proposed methodology is initialized by a k-Means clustering, then the prescribed clusters inform weights of a weighted network, in which players are the nodes and the arcs between them carry those weights that represent a numerical similarity between them. We then call upon a weighted network clustering approach, namely, the Louvain method for community detection. We demonstrate our methodology on six years of historical data, from seasons ranging from 2014–2015 to 2019–2020. Considering these seasons allows us to use a new type of data, called Tracking Data, instated into the league in 2014 which further differentiates our research from other player clustering approaches. We show that our approach can detect outliers and consistently clusters players into groups with identifying features, which give insights into league trends. We conclude that players can be categorized into eight general archetypes and show that these archetypes improve upon the traditional five positions and previous research in terms of explaining variation in Win Shares.