Author(s): CHEN, C., KEMPE, M., EXEL, J., BACA, A., Institution: UNIVERSITY OF VIENNA, Country: AUSTRIA, Abstract-ID: 1960

INTRODUCTION:
Basketball is inherently dualistic often shaped by 1 vs 1 interactions of players. While basketball is on the forefront of sport analytics, match performance indicators don’t account for how and if the player is guarded by a defender. Automatically identifying defensive matchups is therefore a fundamental yet challenging task. Existing approaches are either simplistic, relying on the nearest-defender, or highly complex machine learning models, which achieves high accuracy but at considerable computational cost [1,2]. In this study, we aim to propose a feature-based optimization framework for defensive assignment and assess its performance against ground truth data.
METHODS:
SportVU tracking data from the 2015–2016 NBA regular season was used, providing ball and player positions at 25 Hz. A total of 1,228 possessions from 20 games are randomly selected, ending with shot attempts, standardized to a right-to-left attack, and downsampled to 1 Hz. We use motion-based features commonly applied in man-marking and ball pressures studies. After normalization, a cost function C_ij with initial weights from domain knowledge is built, where i denotes attacker and j defenders. Each frame’s cost matrix is converted to an assignment probability matrix P_ij. Then, feature weights are optimized by minimizing row and column entropy L using differential evolution. To evaluate model performance, an experienced sports analyst annotated defensive assignments for 20 random possessions by watching game footage.

(1) C_ij = w1*I(i,ball-handler) + w2*d(i,basket) + w3*d(i,j) + w4*||v_i - v_j|| + w5*Delta_theta(i,j)
(2) P_ij = exp(-C_ij) / sum_k exp(-C_ik)
(3) L = sum_i H(P_i) + lambda sum_j H(P_j)
H(p) = - sum_k p_k log(p_k)
RESULTS:
Our framework achieves an overall accuracy of 63.67% against expert-annotated ground truth. Among all features, distance between opponents has the largest absolute weight, making it the most influential factor in defensive assignment. Ball handler and velocity vector difference also play substantial roles. Attacker-basket distance contributes moderately, while angular difference has the smallest impact.
CONCLUSION:
Our study proposes an entropy-based soft optimization framework for defensive assignment, achieving 63.67% accuracy against expert labels. The study also shows that the location of the ball, players and the basket are 3 key elements for identifying defensive assignments, which places certain demands on the game dataset. By now accuracy remains below complex machine learning models, while having minimal computing costs. Future work should incorporate richer contextual and temporal constraints to improve robustness.

[1] Keshri S, Oh M, Zhang S and Iyengar G 2019 Automatic event detection in basketball using HMM with energy based defensive assignment J. Quant. Anal. Sports 15 141–53
[2] McQueen A, Wiens J and Guttag J 2014 Automatically recognizing on-ball screens 2014 MIT sloan sports analytics conference