Where Will Players Move Next? Dynamic Graphs and Hierarchical Fusion for Movement Forecasting in Badminton

Kai-Shiang Chang, Wei-Yao Wang, Wen-Chih Peng

TL;DR

This paper introduces the task of movement forecasting in turn-based sports, predicting not only the next shot type in a badminton rally but also where both players will position themselves on the court. The authors present DyMF (Dynamic Graphs and Hierarchical Fusion for Movement Forecasting), a model that structures rallies into relational graphs, captures dynamic tactical changes, and hierarchically fuses rally context with mutual player styles.

What Problem Does This Paper Address?

In fast-paced racket sports like badminton, court positioning is as critical as stroke selection: players must return a shot while anticipating the opponent's reply and moving to an optimal defensive stance. Prior sports analytics research focused largely on stroke forecasting (predicting shot type and where the shuttle lands), neglecting the continuous on-court movement and positioning habits of both competitors.

Tackling this as a sequence prediction task presents distinct challenges:

  1. Relational intent: Returning a shot and returning to base defense have completely different movement semantics (e.g., corner vs. center court).
  2. Dynamic tactics: Players adapt their strategies dynamically within a single rally.
  3. Multifaceted interactions: A player's next move depends on rally progress, their opponent’s immediate response, and mutual playing styles.

Standard sequential models (e.g., LSTMs, Transformers) ignore structural interactions, while existing graph neural networks (GNNs) lack mechanisms to capture turn-based sports dynamics and strategic relationships.

Methods

The authors introduce DyMF, an encoder-decoder architecture built around a custom graph formulation:

  • Player Movements (PM) Graph: A dynamic multi-graph where nodes represent the 2D coordinates of players at each stroke. Edges represent 12 strategic relation types: 10 specific stroke types (smash, lob, clear, net shot, etc.) connecting opposing players across turns, plus two self-transition movement types: defend (moving to defensive position) and return (moving to hit the shuttle).
  • Interaction Style Extractor: Combines two sub-networks:
  • Relational GCN (R-GCN): Propagates information across the 12 relation types using basis decomposition to model structural court interactions.
  • Dynamic GCN: Generates adaptive graph weights using a 1D-CNN + LSTM pipeline over past movement patterns to capture a player’s shifting personal style.
  • Hierarchical Fusion:
  • Player-Player Fusion: Employs parallel co-attention to model how each player's dynamic style influences the other.
  • Player-Rally Fusion: Dynamically balances the relative weight of overall rally flow versus player-specific interaction styles.
  • Movement Predictor: Outputs the next shot type via a linear classifier and predicts the next 2D coordinates for both players by parameterizing a bivariate Gaussian distribution ($\mu_x, \mu_y, \sigma_x, \sigma_y, \rho$).
[PM Graph (Nodes: Player coords, Edges: 12 relation types)]
                         │
        ┌────────────────┴────────────────┐
        ▼                                 ▼
 [Relational GCN]                 [Dynamic GCN]
 (Structural Interactions)        (Pattern-Generated Weights)
        │                                 │
        │                       [Player-Player Fusion]
        │                       (Mutual Style Influence)
        └────────────────┬────────────────┘
                         ▼
               [Player-Rally Fusion]
                         ▼
     [Movement Predictor: Shot Type + Bivariate Gaussian]

Datasets

The authors evaluate on a real-world badminton singles dataset (originally introduced by Wang et al., 2022b):

  • Scale: 75 professional matches, 180 sets, 4,325 rallies, and 43,191 strokes played by 31 high-ranking international players.
  • Data representation: Each stroke entry records player identities, 2D coordinates for both athletes, and 1 of 10 shot types. Rallies are truncated to a maximum length of 35 strokes.
  • Split: Matches are chronologically split per match into the first 80% for training and the final 20% for testing.

Key Insights and Ideas

  • Intent-driven edge semantics matter: Simply modeling players as recurring nodes across time (like dynamic relational inference / dNRI) fails badly. Distinguishing why a player moved (to defend vs. to strike) via specialized graph edges is critical for accurate trajectory and destination modeling.
  • Interactive feedback loop: Badminton tactics are inherently reactive. Counterfactual case studies revealed that altering Player A's simulated defensive positioning forced predictable changes in Player B’s shot selection (e.g., shifting toward the net forced a clear instead of a passive net shot).
  • Failure of continuous dynamic GNNs: Methods like EvolveGCN struggled heavily (producing extreme errors on longer rallies), likely because turn-based rallies have too few nodes per graph snapshot, leading to instability in evolving recurrent graph states.

Key Results

DyMF was evaluated across three observation lengths ($\tau = 2, 4, 8$ prior strokes) against sequence baselines (Seq2Seq, Transformer, ShuttleNet) and graph baselines (dNRI, GCN, R-GCN, EvolveGCN). Performance was measured using MSE and MAE for court location coordinates, and Cross-Entropy (CE) for shot-type prediction.

  • Consistent SOTA: DyMF achieved the best score across every metric and observation length (average rank: 1.0).
  • Quantitative improvements: Relative to competitive baselines, DyMF improved location accuracy by up to 35.3% in MSE and 21.5% in MAE, and shot prediction by up to 24.3% in CE.
  • Representative benchmark ($\tau = 4$):
  • DyMF (Ours): MSE 1.1061 | MAE 1.5888 | CE 1.9487
  • ShuttleNet (2nd best): MSE 1.1539 | MAE 1.6364 | CE 1.9575
  • Seq2Seq: MSE 1.1880 | MAE 1.6739 | CE 1.9905

Limitations and Future Work

  • Singles only: The current formulation is restricted to 1v1 singles play; it does not model the intra-team coordination, space division, and complex rotations inherent to doubles matches.
  • Discretized stroke timing: The model predicts locations conditioned on discrete hit events rather than continuous-time physical velocity, shuttle flight trajectory, or stamina depletion.
  • Future directions: The authors plan to extend the PM graph framework to badminton doubles, where modeling multi-agent interactions and teammate-opponent spatial synergies will be essential.

Summary updated

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. Code is available at https://github.com/wywyWang/CoachAI-Projects/tree/main/Movement%20Forecasting.
BibTeX
@inproceedings{chang2023where,
  author = {Kai-Shiang Chang and Wei-Yao Wang and Wen-Chih Peng},
  title = {{Where Will Players Move Next? Dynamic Graphs and Hierarchical Fusion for Movement Forecasting in Badminton}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  url = {https://openreview.net/forum?id=q6ddstf4c9},
}

Loading related content...

Figure