Efficient Figure Skating Data Analysis using Multi-Task Graph Attention Networks

Figure skating scoring relies on human judgment, leading to subjectivity and bias, highlighting the need for an objective scoring system. In this study, we propose a multi-task graph attention network to simultaneously classify figure skating movements into four tasks: major action category, sub-action category, skill success/ failure, and skater skill level. We constructed a heterogeneous directed skeleton graph which consist of joint nodes with (x, y) coordinates and two types of edges, bidirectional spatial edges between each joints and unidirectional temporal edges to represent the temporal flow of the same joint. Using only lower-body data, our model achieved the best cross-entropy loss, accuracy, and efficiency in terms of time complexity, data utilization, and task performance. In future work, we aim to extend the model for real-time evaluation in a real-world competition environment.