Multi-agent deep-learning based comparative analysis in basketball

抄録

Analysis of multi-agent trajectories is one of the fundamental issues for understanding real-world biological movements. For trajectory analysis, combining with labels (e.g. scored or not in ballgames) can obtain insights rather than only from trajectories. However, the previous deep-learning based method used only single agent trajectory in animals and cannot be directly applied to multi-agent ballgame trajectories. In this paper, we propose a comparative analysis method to analyze multi-agent trajectories in basketball. We adopt a neural network approach using multi-agent motion characteristics (e.g., distances between agents and objects) as the input and based on an attention mechanism to automatically detect segments in trajectories that are characteristic of one group. It enables us to understand the difference between groups by highlighting segmented trajectories and which variables correlate with the labels. We verified our approach by comparing various baselines and demonstrated the effectiveness of our method through use cases that analyze the attacking plays in the NBA league data.