MoCo-AIS: Contrastive Vessel Trajectory Similarity Framework

MoCo-AIS is a unified contrastive learning framework for vessel trajectory similarity that was submitted to arXiv on 16 Jun 2026 (arXiv:2606.17978) by Ruixin Song, Md Mahbub Alam, Zahra Sadeghi, Amilcar Soares, José F. Rodrigues-Jr and Gabriel Spadon. The paper frames similarity learning via the Momentum Contrast (MoCo) paradigm and evaluates a diverse set of leading deep learning models on large-scale, real-world vessel-tracking AIS datasets.

The authors position MoCo-AIS against two limits they see in prior work: traditional distance-based similarity measures, which they say incur high computational cost, and supervised learning approaches, which rely on extensive labels derived from those same distance measures and therefore tend to reproduce their biases and limit generalization. MoCo-AIS adopts self-supervised contrastive learning to create trajectory embeddings and, the paper argues, offers a single benchmarking platform for comparing representation models.

How does MoCo-AIS work and how was it tested?

MoCo-AIS builds on the Momentum Contrast paradigm, formulating similarity learning through positive and negative trajectory pairs, and trains embeddings for vessel trajectories; the authors applied this unified framework to evaluate multiple leading deep learning models on large-scale, real-world vessel-tracking AIS datasets. The paper uses MoCo-style contrastive objectives to avoid explicit label dependence and to unify evaluation across models.

The submission describes a contrastive pipeline that treats pairs of trajectories as positives or negatives and uses momentum encoders to stabilize representation learning. The authors contrast this self-supervised approach with supervised pipelines that depend on labels produced by classical distance metrics, arguing that supervised models often reproduce those metrics rather than generalize beyond them. MoCo-AIS therefore serves two roles in the study: a learning framework and a benchmarking platform for trajectory representation models.

Evaluation, as reported in the abstract, covered "a diverse set of leading DL models" and used "large-scale, real-world vessel-tracking AIS datasets" that capture varied navigation behaviors and operating conditions. Results, the paper states, show that the framework "significantly improves similarity learning over existing baselines," and also establishes a standard way to compare deep models for trajectory representation.

Why does this matter?

MoCo-AIS addresses a practical bottleneck in mobility analysis: computing and comparing trajectory similarity at scale. If contrastive embeddings can replace repeated distance calculations and reduce dependence on labels derived from those same distances, practitioners could run route pattern extraction, mobility prediction and anomaly detection with lighter runtime costs and with representations that generalize beyond classical metrics. The paper’s emphasis on a unified benchmarking platform also matters because it gives researchers a common yardstick to compare models trained under self-supervision rather than bespoke supervised setups.

What to watch

The paper was submitted on 16 Jun 2026 and is listed as under review at SIGSPATIAL'26; the conference review outcome will be the immediate milestone to follow. The arXiv identifier is arXiv:2606.17978 and the author list begins with Ruixin Song and ends with Gabriel Spadon.