Position based clustering in basketball with machine learning

In modern basketball, although the traditional five-position classification system forms the foundation of the game structure, it has become insufficient to comprehensively characterize the growing diversity of players within positions and the variety of playing styles that have evolved over time. This thesis proposes a position-based clustering method that analyzes players using advanced statistical metrics and five-man lineup data, while considering the traditional positions of point guard (PG), shooting guard (SG), small forward (SF), power forward (PF), and center (C). By employing the Bisecting K-means algorithm, which has been relatively underexplored in the NBA analytics literature, player clusters were formed for each position based on normalized advanced statistics covering 15 seasons from 2007 to 2024. To determine the number of clusters, 28 different cluster validity indices were jointly evaluated to identify the most optimal cluster configurations. Beyond clustering players, the study also examines how these clusters influence five-man lineup metrics, thereby revealing how different player types contribute to team strategies. Furthermore, the effectiveness of the Bisecting K-means algorithm is compared with the more commonly used Hierarchical clustering method. This thesis contributes to the field of sports analytics by presenting an interpretable and applicable position-based clustering approach aimed at identifying diverse player profiles and better understanding their impact on team composition.