We model daily global port interactions from AIS origin–destination data (2013–2024) as a dynamic directed graph. Nodes represent ports with static and temporal attributes (e.g., rolling calls and DWT aggregates, capacity proxies, calendar and exogenous signals), while edges capture day-specific traffic features (trip counts, mean speed, vessel mix, delays). We propose a spatiotemporal architecture that applies graph attention networks (GAT) to each weekly snapshot to learn relational embeddings, followed by a gated recurrent unit (GRU) to capture temporal dependencies across a window of past days. This approach leverages network effects beyond handcrafted metrics such as Page Rank or betweenness, enabling the model to learn which neighboring ports and edges most influence a port’s throughput; attention coefficients provide interpretability for scenario analysis. We further demonstrate counter factual simulation by editing graph structure and features (e.g., temporary port closures, speed regulation, new services), yielding predicted impacts on DWT and revealing rerouting patterns.
Link to the presentation (Unzip and open webinar_presentation.html)
Research is forthcoming.
