Similar Sports Play Retrieval with Deep Reinforcement Learning
Zheng Wang,Cheng Long,Gao Cong
DOI: https://doi.org/10.1109/tkde.2021.3136881
IF: 9.235
2021-01-01
IEEE Transactions on Knowledge and Data Engineering
Abstract:With the proliferation of commercial tracking systems, sports data is being generated at an unprecedented speed and the interest in sports play retrieval has grown dramatically as well, where a play corresponds to a fragment of a game. Existing solutions for similar play retrieval usually assume that a database of plays are materialized, which, however, is not well aligned with the practice that data is stored in units of game. In this paper, we propose to search for similar plays directly from a database of games. We tackle three challenges of the task, namely (1) how to measure the similarity between two plays, (2) how to efficiently find a similar play to a query play within a game, and (3) how to efficiently find a similar play within a database of many games. For the first challenge, we propose a deep learning approach called play2vec to learn the representations of sports plays. play2vec is robust against noise and runs in linear time. For the second challenge, we develop a suite of algorithms including two based on reinforcement learning, which use learned policies for deciding where to split a game to generate candidate plays. For the third challenge, we develop a method called ScoreSearch based on deep metric learning, which is able to prune games from being searched for better efficiency. We conduct experiments on real-world soccer match data to evaluate the techniques developed in this paper.
computer science, information systems, artificial intelligence,engineering, electrical & electronic