Exploitation of Hidden Context in Dynamic Movement Forecasting: A Neural Network Journey from Recurrent to Graph Neural Networks and General Purpose Transformers
Schelenz L, Rajanna S, Gosalci D, Heublein L, Pirkl J, Ott J, Ott F, Feigl T (2025)
Publication Language: English
Publication Type: Conference contribution, Original article
Publication year: 2025
Pages Range: 1-26
Conference Proceedings Title: Proc. Intl. Conf. IEEE Symposium on Position Location and Navigation (PLANS)
Event location: Salt Lake City, Utah
Abstract
Forecasting information from past measurements in signal processing pipelines helps mitigate delays caused by the pipeline. Predicting the movement of dynamic objects, such as NBA players, poses unique challenges due to its highly interactive and unpredictable nature. Accurate forecasting of nonlinear movements, such as player positions, requires understanding both individual behaviors and team dynamics. Traditional methods, including statistical and filtering techniques, struggle with the complex, nonlinear nature of sports movements. In contrast, recurrent neural networks, particularly LSTMs, excel by leveraging historical data and modeling player interactions, yielding more accurate trajectory forecasts. However, challenges remain in understanding how hidden contextual information in team sports impacts forecasting accuracy and robustness.
This paper provides a comparative analysis of techniques, including Graph Attention Networks (GAT), Transformer models, Temporal Convolutional Neural Networks (TCNN), and recurrent neural networks like LSTMs, in the context of NBA player position forecasting. Experiments evaluate these methods based on input history length, generalizability, uncertainty, computational complexity, and their ability to exploit implicit contextual team dynamics. Results indicate that time-sensitive AI methods outperform traditional models at all forecast horizons. Among these, carefully trained LSTMs, with explicitly modeled contextual information, achieve superior position accuracy, robustness, and reduced computational effort compared to GAT, Transformer, and TCNN models. Additionally, forecasting future positions effectively compensates for delays inherent to signal processing pipelines.
Authors with CRIS profile
Involved external institutions
Germany (DE)How to cite
APA:
Schelenz, L., Rajanna, S., Gosalci, D., Heublein, L., Pirkl, J., Ott, J.,... Feigl, T. (2025). Exploitation of Hidden Context in Dynamic Movement Forecasting: A Neural Network Journey from Recurrent to Graph Neural Networks and General Purpose Transformers. In Proc. Intl. Conf. IEEE Symposium on Position Location and Navigation (PLANS) (pp. 1-26). Salt Lake City, Utah.
MLA:
Schelenz, Lukas, et al. "Exploitation of Hidden Context in Dynamic Movement Forecasting: A Neural Network Journey from Recurrent to Graph Neural Networks and General Purpose Transformers." Proceedings of the IEEE/ION Position, Location and Navigation Symposium, Salt Lake City, Utah 2025. 1-26.
BibTeX: Download