A Novel Methodology for Automating Spatio-Temporal Data Classification in Basketball Using Active Learning

The use of machine learning on spatio-temporal datasets has generated significant interest in a range of applications, including vehicular traffic modelling and urban planning. One of the most prolific application domains is sports analytics due to the availability of real-world multi-agent datasets, where such techniques are used to recognize and predict offensive and defensive strategies in a range of team sports. However, the use of advanced machine learning techniques requires the large datasets to be annotated by domain experts, which is a time-consuming task. Active learning is a methodology that significantly cuts down the data-annotation time on large datasets. In this paper, we investigate active learning strategies to annotate spatio-temporal datasets for the purpose of classification model building. The proposed algorithms are demonstrated on a dataset obtained from professional basketball games to classify an offensive strategy known as ‘Pick-and-Roll’. Several neural network architectures are investigated for the classification of more than 900 segments of basketball plays. The results obtained suggest that the proposed, preferred, methodology is well suited for annotating large spatio-temporal datasets and has the potential to be applicable across a range of team sports and non-sports usage scenarios.