Abstract: Sports analytics has captured increasing attention since analysis of the various data enables insights for training strategies, player evaluation, etc. In this paper, we focus on predicting what types of returning strokes will be made, and where players will move to based on previous strokes. As this problem has not been addressed to date, movement forecasting can be tackled through sequence-based and graph-based models by formulating as a sequence prediction task. However, existing sequence-based models neglect the effects of interactions between players, and graph-based models still suffer from multifaceted perspectives on the next movement. Moreover, there is no existing work on representing strategic relations among players' shot types and movements. To address these challenges, we first introduce the procedure of the Player Movements (PM) graph to exploit the structural movements of players with strategic relations. Based on the PM graph, we propose a novel Dynamic Graphs and Hierarchical Fusion for Movement Forecasting model (DyMF) with interaction style extractors to capture the mutual interactions of players themselves and between both players within a rally, and dynamic players' tactics across time. In addition, hierarchical fusion modules are designed to incorporate the style influence of both players and rally interactions. Extensive experiments show that our model empirically outperforms both sequence- and graph-based methods and demonstrate the practical usage of movement forecasting.
Introduction
In high-speed sports like badminton, winning is a product of strategic positioning and shot selection. While sports analytics has traditionally focused on predicting shot types (e.g., smash vs. drop), forecasting player movement remains a significant challenge. Predicting where a player will move to recover or intercept the shuttle is crucial for tactical analysis and player evaluation. This paper introduces DyMF (Dynamic Graphs and Hierarchical Fusion), a novel deep learning framework designed to forecast both the next shot type and future court positions by representing rallies as dynamic graphs.
The Challenge of Movement Forecasting
Forecasting movement in badminton involves three primary difficulties. First, representing turn-based, two-player interactions as a structured graph is non-trivial. Second, player tactics are dynamic; they adapt mid-rally based on the opponent's behavior. Third, a player's next move is influenced by a complex interplay of their own habits, their opponent's style, and the current state of the rally.
To address these, the authors introduce the Player Movements (PM) graph. Unlike simple coordinate sequences, the PM graph treats nodes as player locations and edges as "strategic relations." These relations include 10 standardized shot types and two movement intents: Defend (moving to a recovery position) and Return (moving to intercept).
Methodology: The DyMF Framework
The proposed DyMF model employs an encoder-decoder architecture designed to capture the nuances of player interactions. It utilizes the PM graph to exploit structural movements and strategic relations.
Key components include:
- Interaction Style Extractor: This module combines Relational GCNs to handle diverse edge types with Dynamic GCNs (using 1D convolutions and LSTMs) to extract evolving movement patterns over the course of a rally.
- Hierarchical Fusion: This mechanism operates in two stages. Player-Player Fusion uses co-attention to model mutual influence between opponents. Player-Rally Fusion then integrates these individual styles with the broader context of the rally progress.
The model outputs a dual prediction: a classification for the next shot type and a bivariate Gaussian distribution for the (x, y) coordinates of both players.
Experimental Evaluation
The study utilized a real-world dataset of professional badminton singles, comprising 75 matches, 31 players, and over 43,000 strokes. DyMF was benchmarked against standard sequence models (Seq2Seq, Transformers) and specialized sports models like ShuttleNet.
Key Results
- Superior Accuracy: DyMF significantly outperformed baselines, achieving up to a 35.3% improvement in Mean Square Error (MSE) for location prediction and a 24.3% improvement in Cross-Entropy (CE) for shot type prediction.
- Dynamic Modeling: Ablation studies confirmed that modeling the interaction and the dynamic nature of tactics is essential; removing these components led to substantial performance degradation.
- Intent-Based Insights: By explicitly labeling movement as "Defend" or "Return," the model successfully learned that recovery movements are functionally distinct from interceptive lunges.
Conclusion
DyMF represents a significant step forward in sports trajectory forecasting by treating movement as a functional, strategic interaction rather than just a spatial sequence. While currently focused on singles, future work aims to extend this to doubles, accounting for the more complex coordination between partners. The model's ability to predict distributions also opens doors for counterfactual "what-if" analysis in coaching and strategy development.
Summary updated