講演情報
[4O1-IS-2a-06]Dual-Stream Soccer Event Prediction via Disentangled Trajectory Representations
〇Zheng CHEN1, Calvin Yeung1, Keisuke FUJII1,2 (1. Nagoya University, 2. Institute of Physical and Chemical Research (RIKEN))
work-in-progress,[[online]]
キーワード：
Multi-agent systems、Soccer analytics、Representation learning、Trajectory prediction、Event prediction
Soccer is a complex multi-agent system characterized by coupled tactical intents and physical movements. However, existing decoupled approaches treat these components as disjoint outputs, resulting in kinematic inconsistency and a failure to encapsulate a holistic event prediction. Here we propose a unified dual-stream framework that strictly enforces kinematic anchoring, ensuring every predicted event is grounded in our predicted trajectories. A fundamental challenge in such coupled modeling is that the inherent uncertainty of trajectory prediction introduces noise and error accumulation that compromises semantic inference. To address this, we introduce an event-guided disentangled representation learning mechanism. By orthogonally factorizing motion features to distill stable semantic intent, our model selectively prioritizes robust tactical signals while suppressing noisy dynamic fluctuations. To further facilitate the optimization of these coupled objectives, we employ a curriculum learning strategy that progressively aligns trajectory dynamics with event semantics. Experiments on a large-scale LaLiga dataset demonstrate that our approach achieves superior performance compared to baselines. By effectively mitigating the negative impact of trajectory noise, our framework offers an effective approach for holistic and consistent modeling in soccer data.