Analyzing the Relationship Between Football Players’ Passing Network Structure and Tactical Performance Using Graph Convolutional Networks

Current football passing-network analysis methods lack dynamic and interpretable modeling of the relationship between network structure and tactical performance. Based on the GCN framework, this paper constructs a T-GAT model by integrating a dynamic attention mechanism and GRU temporal modeling. The match is represented as a directed attributed passing graph over time windows. An edge-aware multi-head GAT encoder generates window-level graph representations, which are then fed into a GRU with time decay and retrospective attention to form hybrid temporal representations under unequal time intervals. The model jointly predicts window-level ∆xG, ∆xT, Progressive Passes, and Pass Network Penetration Rate, while extracting edge attention for visualization and interpretation of key passing links. Experimental results show that T-GAT achieves ∆xG RMSE values of only 0.06 and 0.09 in the 0–15 and 76 –90 minute test periods, respectively. Structural correlation analysis indicates that passing-network density is strongly correlated with ∆xG (Pearson r = 0.84, p < 0.001), while average node output strength is correlated with ∆xT (r = 0.92, p < 0.001). The findings show that high connectivity and decentralized control improve tactical performance and provide reference for networked propagation and signal-flow analysis.