Two-Stage Recommendation System Solves the Ranking Challenge of Tropical Cyclone Port Impacts

When a tropical cyclone approaches, port authorities care about two questions above all: which ports will be affected, and where will disruptions last longest? A recent study proposes a two-stage recommendation framework to answer these questions—first locking onto candidate ports near the landfall point, then ranking them by estimated impact days. In tests with Hurricanes Nicole and Ian, the method's ranking consistency was 85.5% and 55.5% higher than a distance baseline.

The Problem: Why Port Impact Estimation Is Difficult

When tropical cyclones threaten multiple ports, disaster preparedness planning depends on anticipating which ports may be affected and how long disruptions may last. But two real-world obstacles make estimating relative impacts tricky: historical observational data for any single port is limited, and the landfall point itself is uncertain. Traditional approaches tend to rely simply on the distance between a port and the forecast landfall point, but proximity does not equal longer impact duration, which prompted researchers to turn to more refined modeling approaches.

The Method: A Two-Stage Framework of Retrieval Then Ranking

The study adopts a two-stage recommendation framework. The first stage is retrieval, filtering candidate ports based on their proximity to the expected landfall point; the second stage is ranking, ordering candidate ports by estimated impact days.

The study region is the Gulf of Mexico, with data covering 37 ports and 31 tropical cyclones, where impact days are derived from AIS (Automatic Identification System) data. For representing port-cyclone interactions, the researchers compared three approaches: pairwise feature interactions, tree-based feature combinations, and explicit feature crosses combined with learned embedding representations.

Wide & Deep learning models are a recommendation system architecture proposed by Google in 2016, jointly trained from a "Wide" part and a "Deep" part. The Wide part is a generalized linear model that excels at memorizing frequently occurring feature-cross combinations in the training data (such as "specific port × specific cyclone intensity"), but has limited generalization ability; the Deep part is a multi-layer feedforward neural network that maps sparse high-dimensional features into low-dimensional dense vectors through an embedding layer, enabling generalization to unseen feature combinations. The outputs of the two parts are summed and passed through a Sigmoid to produce the final prediction. In this study, the advantage of Wide & Deep is that it can capture strong prior linear relationships such as "port–landfall distance" while also learning nonlinear interactions among port features, cyclone features, and impact days through the deep part, thereby maintaining ranking stability even when the landfall point shifts.

AIS (Automatic Identification System) is a broadcast system that the International Maritime Organization mandates for ships. Vessels periodically transmit static and dynamic information—including their position, speed, heading, and draft—over VHF channels, and shore-based receiving stations and satellites can receive these signals. In port impact assessment, AIS data can reflect the actual level of vessel activity around a port: when disruptions occur, the number of vessels entering and leaving the port drops sharply and vessels pile up at anchorages, and these anomalous patterns can be used to infer the start, end, and duration of port disruptions. Compared with official disruption records from port authorities, AIS data has broad coverage, high temporal resolution, and is not affected by inconsistent post-hoc reporting standards, making it the data source for deriving the key label of "impact days" in this study.

Results: Wide & Deep Model Rankings Closest to Reality

In tests with Hurricanes Nicole and Ian, the retrieval stage consistently identified all actually affected ports under landfall uncertainty, showing that this stage can effectively avoid false negatives.

The ranking stage showed clear differences in performance. The Wide & Deep Learning model had the highest consistency with the actually observed port rankings: its average ranking consistency was 85.5% higher than the distance baseline for Nicole and 55.5% higher for Ian, and it was less sensitive to uncertainty in the landfall location. This means that even if the forecast landfall point shifts, the port priority ranking produced by the model remains relatively stable.

Ranking consistency is a metric in recommendation systems and information retrieval for measuring ranking quality, commonly quantified using Kendall's tau or Spearman's rank correlation coefficient. The basic idea is to compare the degree to which the pairwise orderings in the model's port ranking agree with the true observed ranking—if the model ranks port A before B and the actual impact days are also longer for A than for B, then this pair is consistent; otherwise it is inconsistent. The higher the consistency, the better the model can correctly distinguish which ports are more severely affected. This study uses the distance baseline as a reference and reports the percentage improvement of the model over the baseline to illustrate the actual gain in ranking quality from incorporating AIS impact days and nonlinear feature interactions.

Application Value: Resource Allocation and Post-Disaster Inspections

The direct use of such ranking results is to help port authorities make trade-offs when multiple ports are threatened simultaneously—which ports' post-disaster inspections to prioritize, and how to allocate recovery resources across ports. Compared with coarse judgments based only on distance, rankings based on AIS impact days are closer to actual disruption durations, helping direct limited emergency resources to where they are truly needed.

Limitations and Outlook

The study is based on 37 ports and 31 cyclone samples in the Gulf of Mexico, so the sample size is limited; the model's generalization ability in other sea areas and under different cyclone tracks still needs verification. In addition, although landfall uncertainty is taken into account, how changes in forecast error itself affect ranking stability is still worth further exploration.

For port operations and maritime emergency management, the insight of this work is that introducing the retrieval-ranking idea from recommendation systems into disaster impact assessment can provide an actionable port priority list under conditions of limited information, without having to wait for precise physical impact simulations.

Landfall uncertainty stems from the inherent error in tropical cyclone track forecasts. Meteorological agencies typically use a "cone of uncertainty" to represent the probability range of possible landfall locations; the longer the forecast lead time, the wider the cone, and the standard deviation of the landfall location can reach tens to over a hundred kilometers. For port impact assessment, this means the retrieval stage cannot simply lock onto ports near a single landfall point but must cover all candidate ports within the possible landfall area; otherwise, false negatives may occur. This study incorporates landfall uncertainty into the design of retrieval and ranking precisely to avoid missing actually affected ports due to forecast shifts; and the ranking model's insensitivity to this uncertainty means that even if the landfall forecast deviates, the port priority list can still remain relatively reliable.

One-line Summary

A two-stage recommendation framework retrieves then ranks to estimate port disruption days under tropical cyclones, with ranking consistency significantly better than a distance baseline.

Paragraph Summary

The study addresses the difficulty of estimating port impacts when tropical cyclones approach—limited historical data for individual ports and landfall uncertainty—by proposing a two-stage recommendation framework: first retrieving candidate ports based on proximity to the expected landfall point, then ranking them by estimated impact days derived from AIS data. Using data on 37 ports and 31 cyclones in the Gulf of Mexico, it compared three representation approaches: pairwise feature interactions, tree models, and explicit feature crosses plus embeddings. Tests with Hurricanes Nicole and Ian showed that the retrieval stage could still identify all actually affected ports under landfall uncertainty; the Wide & Deep model's ranking consistency was 85.5% and 55.5% higher than the distance baseline, respectively, and was more robust to landfall shifts. This ranking can help port authorities prioritize post-disaster inspections and cross-port resource allocation. The study sample is limited, and generalization ability and the impact of forecast error still need verification.