Ryan Gunther, a Master of Science in Statistics candidate, will defend his thesis titled Dimensionality Reduction Approaches for Baseball Swing Motion Analysis on Wednesday, August 12, 2026 at 11:00 AM in GSB 405.
The examination committee includes Melanie Pilkington, Chair; Shawn Beaudette, External Examiner; William Marshall, Supervisor; Jan Vrbik and S. Ejaz Ahmed, Supervisory Committee Members.
Abstract:
Background: Baseball biomechanical data are often high-dimensional, temporally structured and difficult to analyze directly. Although baseball biomechanics research has been heavily invested into over the past 10 years, the majority of it has been to the benefit of pitchers rather than batters. This thesis aims to investigate whether dimensionality reduction methods for large baseball biomechanics datasets of batters could successfully create compact representations of swing motion while preserving information relevant to bat speed.
Methods: First, dimensionality reduction techniques (principal-feature selection, probabilistic principal component analysis, covariance-based temporal segmentation, and clustering) were performed on a public biomechanics dataset of 431 marker-based motion capture swings, reducing them to feature vectors to perform cross-validated LASSO regression on. Separately, a computer vision pipeline and corresponding 200 swing MLB dataset were built. Hitter silhouettes were segmented, transformed into distance-weighted images, and embedded as two-dimensional Isomap trajectories, again with the goal of extracting meaningful features from a reduced dataset. Ridge regression and recurrent neural networks were trained on the data (simple RNN, GRU, and LSTM) to examine predictive performance.
Results: Both datasets were successfully reduced to a compact representation that could be used for visualization, comparison, and downstream modelling, but predictive gains were limited. Across the marker-based and markerless datasets, models produced little improvement over the mean baselines. Adding first- and second- derivative features and adjusting bat speed for pitch context did not materially change this result.
Conclusion: The representations created a bridge between raw motion data and interpretable analysis, allowing swing movement to be studied in forms that would be difficult to access from the original high-dimensional inputs alone. In the available datasets, the reduced representations did not support strong out-of-sample prediction of bat speed (or exit velocity), but they provide a framework for organizing complex motion data for future analysis.
