Decoding Defensive Coverage Responsibilities in American Football Using Factorized Attention Based Transformer Models

Kevin Song; Evan Diewald; Ornob Siddiquee; Chris Boomhower; Keegan Abdoo; Mike Band; Amy Lee

cs.LGMar 26, 2026

Decoding Defensive Coverage Responsibilities in American Football Using Factorized Attention Based Transformer Models

Authors: Kevin Song, Evan Diewald, Ornob Siddiquee, Chris Boomhower, Keegan Abdoo, Mike Band, Amy Lee

Organizations: Amazon Web Services, Seattle, WA, USA · National Football League, New York, NY, USA

Abstract

Defensive coverage schemes in the National Football League (NFL) represent complex tactical patterns requiring coordinated assignments among defenders who must react dynamically to the offense's passing concept. This paper presents a factorized attention-based transformer model applied to NFL multi-agent play tracking data to predict individual coverage assignments, receiver-defender matchups, and the targeted defender on every pass play. Unlike previous approaches that focus on post-hoc coverage classification at the team level, our model enables predictive modeling of individual player assignments and matchup dynamics throughout the play. The factorized attention mechanism separates temporal and agent dimensions, allowing independent modeling of player movement patterns and inter-player relationships. Trained on randomly truncated trajectories, the model generates frame-by-frame predictions that capture how defensive responsibilities evolve from pre-snap through pass arrival. Our models achieve approximately 89%+ accuracy for all tasks, with true accuracy potentially higher given annotation ambiguity in the ground truth labels. These outputs also enable novel derivative metrics, including disguise rate and double coverage rate, which enable enhanced storytelling in TV broadcasts as well as provide actionable insights for team strategy development and player evaluation.

Figures & tables

Figure 1: Model architecture of the three coverage responsibility models. (A) General architecture that’s shared amongst the 3 models (B) Different output heads for each model.

Time frame	Accuracy	F1
-30 to pass arrival	0.894	0.894
-20 to pass arrival	0.894	0.894
-10 to pass arrival	0.894	0.894
Snap to pass arrival	0.893	0.893
-20 to pass forward	0.873	0.872
-10 to pass forward	0.873	0.872

Table 1: Receiver-defender matchup accuracy across different time frames.

Time frame	Accuracy	F1
-30 to pass arrival	0.918	0.879
-20 to pass arrival	0.918	0.879
-10 to pass arrival	0.918	0.878
-10 to pass forward	0.919	0.879
-20 to pass forward	0.918	0.878
-30 to pass forward	0.918	0.878

Table 2: Individual coverage assignment accuracy across different time frames.

Implementation	Accuracy
Rules-Based Nearest Defender	0.764
Transformer	0.862
Transformer + Post-Processing Rules	0.882

Table 3: Target defender classification accuracy against rule-based method and with and without post-processing rules.

Figure 2: Receiver-defender matchup results over time. Receiver-defender matchup predictions at (A) pre-snap, (B) mid-play, and (C) pass arrival.

Figure 3: Coverage assignment prediction over time. From left to right, the frames show snapshots of predictions (A) 3 seconds before the snap and at (B) pass arrival.

Figure 4: Misclassified individual assignments due to potential mislabeled samples (HIGH_SAFETY1) and/or genuine ambiguity in man vs. zone assignments (CB1). (A) Coverage assignment predictions at pre-snap and (B) coverage assignments at full play both show misclassified predictions. This example shows the model’s bias toward symmetry and respecting the overall coverage scheme.

Figure 5: Misclassified individual assignments due to disagreement on overall coverage scheme, leading to a cascade of predictions that do not align with the human labels, but are internally consistent within the context of the predicted Cover 2 shell.

Figure 6: Play level target defender prediction illustrated at pass forward.

Figure 7: Disguise rate was approximated by calculating pre-snap coverage assignment accuracy and full accuracy.

Figure 8: Comparison of Double coverage rate per player over the 2024 season between the prediction and ground truth.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Training Task	Target Defender	Receiver-Defender Matchups	Coverage Assignments
Batch Size	16	16	16
Num Epochs	100	200	150
Learning Rate and Scheduler	OneCycleLR ( $max\_lr=2e^{-4}$ , $total\_steps=\newline train\_len*100$ , $pct\_start=0.1,\newline div\_factor=10.0$ , $final\_div\_factor=\newline 100.0$ )	OneCycleLR ( $max\_lr=2e^{-4}$ , $total\_steps=\newline train\_len*200$ , $pct\_start=0.1,\newline div\_factor=10.0$ , $final\_div\_factor=\newline 100.0$ )	Cosine Annealing Warm Restarts $init\_lr=2e^{-5}$ $T_{0}=15$ $T_{m}ult=1$ $eta\_min=2e^{-7}$
Optimizer	AdamW weight decay=1e-4, betas=(0.9, 0.999)	AdamW weight decay=1e-4, betas=(0.9, 0.999)	AdamW weight decay=1e-4, betas=(0.9, 0.999)
Num Heads	4	4	8
Num Layers	3	3	6

Appendix

Table 4: Training Optimization and Hyperparameter settings

Explore similar work

Evaluating passing decision-making in professional football: An enhanced MPNN approach to Receiver Selection
May 25, 2026Gabriel Masella, Giuseppe Alessio D'Inverno, Max Goldsmith +1Graph Neural Networks Message Passing Neural Networks
Department of Mathematics, Informatics and Geoscience, University of Trieste, via Weiss 2, Trieste, 34128, Italy. · MathLab, International School for Advanced Studies (SISSA), Via Bonomea 265, Trieste, 34136, Italy. · Royal Belgium Football Association, Rue de Bruxelles 480, Tubize, 1480, Belgium.
Hierarchical Possession-Aware Graph Pointer Network for Pass Receiver Selection
Sep 4, 2026Jingyi Wang, Da Li, Kaixin Wang +1Graph Neural Networks Sports Video Analysis
Beijing Engineering Research Center for IoT Software and Systems, Beijing University of Technology, Beijing, China · Beijing Key Laboratory of Interdisciplinary Intelligent Technologies in Sports Medicine and Engineering, Beijing University of Technology, Beijing, China
Blame is easier than praise: Measuring off-ball defensive performance in football
Jun 18, 2026Jonas Bischofberger, Runqing Ma, Pascal Bauer +2Automated Evaluation Fine-Grained Credit Assignment
Centre for Sport Science and University Sports, University of Vienna, Vienna, Austria · Vienna Doctoral School of Pharmaceutical, Nutritional and Sport Sciences (VDS-PhaNuSpo), University of Vienna, Vienna, Austria · VfB Stuttgart 1893 AG, Stuttgart, Germany +2

May 25, 2026J/K move · Enter open · S save

Evaluating passing decision-making in professional football: An enhanced MPNN approach to Receiver Selection

Gabriel Masella, Giuseppe Alessio D'Inverno, Max Goldsmith, Gianluigi Rozza

Department of Mathematics, Informatics and Geoscience, University of Trieste, via Weiss 2, Trieste, 34128, Italy. · MathLab, International School for Advanced Studies (SISSA), Via Bonomea 265, Trieste, 34136, Italy. · Royal Belgium Football Association, Rue de Bruxelles 480, Tubize, 1480, Belgium.

Graph Neural Networks Message Passing Neural Networks

Open paper AskPDF arXiv

The process of decision-making in football is characterized by a complex interplay between spatial positioning, opponent pressure, and player intent. This work introduces a Graph Neural Network (GNN) framework designed to predict Receiver Selection, the optimal passing target, by modeling on-field interactions as dynamic graphs. Each player is represented as a node with positional and contextual features, while potential passing lines form weighted edges characterized by distance, angle, and pressure metrics. A Message-Passing Neural Network (MPNN) has been developed and trained using a combination of tracking data and event data from professional matches, synchronized through a robust pipeline based on an optimized version of the Needleman-Wunsch Algorithm. The model achieves competitive accuracy in identifying the actual chosen receiver and state-of-the-art accuracy within its top three suggestions. Our model further offers quantification of each option's likelihood, threat, and creativity, enabling performance analysts to evaluate over 1,000 passes in seconds.

Abstract

Explore similar work

Evaluating passing decision-making in professional football: An enhanced MPNN approach to Receiver Selection

Hierarchical Possession-Aware Graph Pointer Network for Pass Receiver Selection

Blame is easier than praise: Measuring off-ball defensive performance in football