TrajTok: Adaptive Spatial Tokenization for Trajectory Representation Learning
Researchers have developed TrajTok, a novel method for learning generalizable trajectory representations from GPS data. The system addresses challenges posed by continuous, noisy, and irregularly sampled data by employing a multi-resolution hexagonal cell partition to convert GPS sequences into discrete tokens. TrajTok utilizes a factorized transformer encoder with specialized attention layers and spatiotemporal rotary position embeddings to encode both the location and timing of these tokens. Pretrained using a masked-token modeling approach, TrajTok demonstrates strong performance across various trajectory-related tasks, suggesting its potential as a general-purpose foundation model for trajectory data. AI
IMPACT Introduces a new foundation model approach for trajectory data, potentially improving downstream applications like similarity search and ETA prediction.