Seeing the forest and the trees: self-supervised trajectory representation for transport dynamics
Ze Wang, Yuan Cheng, Yahong Wang, Ao Li, Fihu Jin · Transportmetrica B Transport Dynamics · 2025
Existing self-supervised contrastive learning methods in trajectory representation learning (TRL) often rely on large negative samples, substantial batch sizes, memory banks, or complex mining strategies and fail to sufficiently capture the temporal patterns in trajectory data. To address this limitation, we propose a novel approach inspired by the Bellman equation – where the current state’s value is estimated based on predicted future states – and by one-dimensional convolution that approximates wavelet transforms for temporal pattern extraction. Our method incorporates a dual-network setup, where the online network learns from augmented views of the same trajectory and captures global structures, while a Masked Language Model (MLM) task is employed to capture local semantic context. Experimental results show that our Attention BERT Contrastive (ABC) framework improves performance by 6.19% on average across tasks like temporal estimation and trajectory classification, demonstrating its effectiveness and scalability for self-supervised TRL.