TrajTok: Adaptive Spatial Tokenization Makes Raw GPS Traces Usable for Foundation Model Pretraining

arXiv·low signal

TrajTok introduces adaptive spatial tokenization that handles the continuous, noisy, irregularly-sampled nature of GPS trajectory data, converting raw traces into discrete tokens suitable for representation learning. This bridges the gap between raw mobility data and foundation model architectures. Relevant for practitioners working with location data, fleet management, or urban computing who want to leverage pretrained representations.