Pairwise Diverse and Uncertain Gradient-Sampling for Similarity Retrieval

Area of Science:

• Computer Science
• Data Science
• Sports Analytics

Background:

• Sports tracking generates vast, unstructured trajectory data crucial for analyzing plays.
• Current similarity search methods for this data rely on dimensionality reduction, often using Siamese Networks.
• Training Siamese Networks is computationally expensive due to the combinatorial nature of pairwise comparisons and distance calculations.

Purpose of the Study:

• To address the computational challenges in training Siamese Networks for sports trajectory data analysis.
• To develop a novel sampling technique that improves the efficiency and effectiveness of representation learning.
• To enhance the quality and speed of similarity search for identifying interesting plays in large sports datasets.

Main Methods:

• Proposed a new sampling technique called Pairwise Diverse and Uncertain Gradient (PairDUG).
• PairDUG leverages model gradient signals to select informative and representative pairs for training.
• Implemented and evaluated PairDUG on large-scale basketball and American football trajectory datasets.

Main Results:

• PairDUG at least halved the required compute time for training.
• Retrieval quality was maintained or improved compared to existing methods.
• Outperformed other baseline sampling techniques in terms of efficiency and retrieval effectiveness.
• Selected pairs via PairDUG demonstrated greater gradient magnitude, diversity, and stability.

Conclusions:

• PairDUG offers a foundational contribution to efficient pairwise distance learning.
• The method significantly improves computational feasibility for training on large trajectory datasets.
• Future work can extend PairDUG to other sports and complex trajectory data domains.