GCN-based learner movement feature network modeling for personalized training path planning and dynamic evaluation

Abstract Personalized training path planning has emerged as an active concern in intelligent sports education, yet many existing approaches struggle to capture the dependencies that link diverse movement features. This study develops a graph convolutional neural network (GCN) approach that models learner movement feature networks for personalized path planning and dynamic evaluation. We build heterogeneous graphs in which motor skills and learner profiles serve as nodes, while edges encode skill correlations and learner similarities through a hybrid weighting strategy. On top of these graphs, a GCN architecture with attention and residual connections learns discriminative embeddings; the path planning module then couples these embeddings with reinforcement learning to assemble training trajectories that respect prerequisite constraints and balance competing objectives. A dynamic evaluation component monitors learner performance and triggers adaptive path adjustments through incremental embedding refinement. On a single-institution university physical education dataset, and averaging five runs that differ only in the random seed, the proposed method reaches 83.67% ± 0.29% path recommendation accuracy and a 24.35% ± 0.38% training effect improvement rate; the margin over the strongest baseline we tested is 4.76% points, with a bootstrap 95% confidence interval of 4.26 to 5.22 points and a paired t-test giving p < 0.001. Dynamic evaluation attains 87.92% ± 0.35% mastery prediction accuracy, against 71.24% ± 0.58% for the static counterpart. The improvement rate comes from a comparison between class sections that were allocated by timetable rather than by randomisation, and is therefore reported as an association. We read these results as evidence that graph-based representations capture useful relational structure in motor skill domains, rather than as proof of broad cross-sport applicability. The scope is deliberately bounded: the framework targets sports organised around hierarchical skill dependencies—gymnastics, swimming, track and field, and structured ball-skill progressions—and the evaluation is in-domain on one institutional corpus rather than a cross-sport generalisation test.