Evaluating defensive performance in football is difficult because much of a defender's value comes from off-ball positioning and from preventing dangerous situations before they occur. This thesis proposes pass suppression value (PSV), a graph-based metric for quantifying how individual defenders reduce the likelihood of dangerous passing options. The method combines two graph attention models applied to SkillCorner tracking and event data from the Premier League 2024/25 season. The first model predicts the likelihood that each potential receiver is selected by the ball carrier, while the second estimates the receiver-conditional threat of each passing option. By masking attention weights between a defender and a potential receiver, the method constructs a counterfactual pass selection probability and measures how much the defender suppresses that option. This suppression is then weighted by the expected threat of the pass.

The pass selection model clearly outperforms an XGBoost baseline, reaching an accuracy of 80.6%, an MRR of 0.890, and Hits@3 of 0.974. The expected threat model provides a smaller improvement over XGBoost, suggesting that relational information is more important for predicting pass targets than for estimating downstream danger. When aggregated over the season, PSV correlates with centre-backs' transfer values, with the strongest correlation obtained when all frames in the possession sequence before the pass are included. The metric also shows promise for distinguishing between centre-backs within the same pairing, an area where traditional defensive statistics are limited. Qualitative examples indicate that PSV can identify meaningful defensive positioning, although some individual outputs remain difficult to interpret. Overall, the results suggest that PSV can support scouting and defensive analysis, while further work is needed to improve interpretability and to extend the evaluation to other aspects of defending.