Unsupervised graph representation learning from AIS maritime traffic

Automatic Identification System (AIS) data provide a large-scale record of maritime vessel movements, yet learning reusable port representations from these routes remains underexplored. This paper investigates whether unsupervised graph representation learning can produce meaningful and operationally useful embeddings for ports in a global maritime port network. We construct a directed attributed port-visit network from 6.15 million AIS-derived vessel journeys and evaluate Deep Graph Infomax (DGI) and related graph embedding methods on structural analysis tasks and next-port prediction. The learned embeddings recover substantial geographic and functional organization, including regional clustering, hub hierarchy, and vessel-type specialization. This demonstrates that unsupervised graph embeddings can capture maritime structure largely from the graph topology. Their predictive utility, however, is strongly task-dependent. While the embeddings support port similarity search and structural benchmarking, they provide limited improvement for next-port prediction, where sequential, frequency-based baselines remain strongest. Further analyses indicate that the high density of the global maritime graph dilutes local neighborhoods, limiting representation separability and reducing the effectiveness of directional message passing. These findings show that unsupervised port embeddings are valuable for structural analysis and peer discovery, but downstream utility depends strongly on the alignment between embedding objectives and the operational structure of the target task.