SPAIT: a novel movement forecasting model with shot-based position-adaptive inference transformer in badminton

Movement forecasting in badminton singles involves predicting shot types and player movements. In practice, the movements of offensive and defensive players correlates with the probability distribution over shot types that a player may execute. Since the movement data of badminton singles is treated as sequences, a natural approach is to utilize sequence-to-sequence models. However, existing sequence-to-sequence models do not take advantage of the correlation between shot types and player movements and rely on restrictive assumptions on player position distributions, whereas graph-based models still struggle with the complexity of possible movements. To address this, we propose a novel movement forecasting model Shot-Based Position-Adaptive Infer Transformer(SPAIT). First, Area Position Feature-Encoder(APF-Encoder) can resolve the issue of transformers handling different data types and fully utilize it by jointly modeling continuous and discrete representations. Second, Guidance Mask(GM) employs masking and positional encoding to recognize feasible movement regions and also accelerate the convergence process. Finally, through the hierarchical derivation characteristics of Infer-Decoder, Infer-Decoder builds meaningful connections between shot types and player positions. This eliminates the failure cases observed in other models, such as predicting player positions inconsistent with the executed shot types. Experimental results show that SPAIT outperforms state-of-the-art movement forecasting models across multiple evaluation metrics for movement forecasting.