Applications of Representation Learning Methods in Professional Baseball

Abstract In the realm of sports analytics, players, teams, and managers have primarily been evaluated through a set of summary counting statistics. Typically, these statistics describe the number of times various events occurred on the field of play, perhaps dividing one quantity by another and/or applying some statistical adjustment. Recent works have proposed viewing the game as a sequence of events towards identifying specific items of interest, such as the players being in a certain formation on the field. However, this sequential, contextual understanding of the game is abandoned once the item of interest is identified. We speculate performance can be gained by leveraging this sequential, contextual understanding of the game. In this work we demonstrate how a rich, event-by-event understanding of a sport, in this case professional baseball, can provide utility in the tasks of player performance projection and clustering. In the process of doing so, we also demonstrate how findings in the domain of sports can translate to state-of-the-art advances in more traditional disciplines of study - specifically, the natural language processing task of emotion recognition in conversation (ERC) and the computer vision task of long-tail partial-label-learning (LT-PLL).