📊 statistics

Vessel Traffic Flow Prediction on Sparse Data via Spatio-Temporal Graph Neural Networks with a Learnable Tweedie Head

This paper proposes a model-agnostic, learnable Tweedie head that optimizes closed-form unit deviance to enhance spatio-temporal graph neural networks, effectively addressing data sparsity and intermittent bursts in maritime traffic flow prediction to improve forecast accuracy for non-zero events.

Original authors: Kyeongjun Lee, Heeyoung Kim

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Kyeongjun Lee, Heeyoung Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a busy port like Los Angeles and Long Beach as a giant, living city of ships. Sometimes, the "streets" (waterways) are packed with vessels moving in and out. Other times, they are eerily empty. This paper tackles the difficult job of predicting exactly when and how many ships will arrive, especially when the data is full of "ghosts" (empty hours) and sudden, chaotic "surges" (many ships arriving at once).

Here is the breakdown of their work, using simple analogies:

The Problem: The "Bored Predictor"

Think of a standard AI model trying to predict ship traffic like a weather forecaster who is terrified of being wrong. Because the data is so often empty (zero ships), this "bored predictor" gets lazy. It decides the safest bet is to just say, "No ships coming," for almost every hour.

The Result: It looks good on paper because it's right most of the time (since there are often no ships). But when a real surge happens—like 50 ships arriving in an hour—it completely misses the boat, predicting zero instead.
The Old Fix: Researchers tried using a "Two-Part" system (called ZINB). Imagine a bouncer at a club who first asks, "Is anyone coming?" (Yes/No) and then, if yes, "How many?" The paper argues this bouncer is too slow and rigid. If the "Yes" gate doesn't open fast enough during a sudden rush, the system still predicts a low number, failing to capture the true size of the crowd.

The Solution: The "Smart, Flexible Head"

The authors built a new tool called a "Learnable Tweedie Head." Think of this as a special, plug-and-play hat you can put on any existing AI model (the "body") to make it smarter about traffic.

Instead of asking a simple "Yes/No" question, this new head understands that ship traffic is like rain:

Sometimes it doesn't rain at all (Zero ships).
Sometimes it drizzles (A few ships).
Sometimes it pours in a sudden, heavy storm (A massive traffic spike).

This "hat" uses a mathematical rule (the Tweedie distribution) that naturally handles this mix of "nothing" and "everything" without needing two separate steps. It learns to say, "Okay, it's usually dry, but when it rains, it can really pour," and adjusts its prediction accordingly.

How They Tested It

They built a digital map of the Los Angeles/Long Beach port using real ship tracking data (AIS). They treated different parts of the port (terminals and anchorages) as "nodes" on a graph, connected by the paths ships take.

They tested their new "hat" on three different types of AI bodies (STGCN, DCRNN, and Graph WaveNet).

The Test: They asked the models to predict the next 6 hours of traffic based on the last 24 hours.
The Metric: They didn't just look at the average error. They specifically looked at how well the models predicted the busy hours (the non-zero spikes), because that's when port managers actually need help.

The Results

The "Bored" Models: Predicted very low numbers, missing the big surges entirely.
The "Two-Part" Models: Got better at seeing that ships were coming, but still underestimated how many were coming during a rush.
The "Tweedie Head" Models: Were the best at catching the surges. They didn't just say "ships are coming"; they predicted the magnitude of the rush much more accurately.

The Bottom Line

The paper claims that by swapping out the standard "final step" of these AI models with their new Tweedie Head, they can stop the models from being "conservative" (playing it safe with zero predictions). This leads to more accurate forecasts when ships actually show up, which is crucial for keeping ports running smoothly and safely.

What they didn't claim:

They did not claim this solves every problem (the model still struggles with the absolute wildest, most extreme spikes).
They did not claim this works for other types of traffic (like cars or planes) yet, only ships in this specific port region.
They did not claim to include outside factors like weather or schedules in this specific version of the model.

Technical Summary: Vessel Traffic Flow Prediction on Sparse Data via Spatio-Temporal Graph Neural Networks with a Learnable Tweedie Head

1. Problem Statement

Accurate vessel traffic flow prediction is essential for smart port operations and navigational safety. However, maritime traffic data presents unique challenges compared to urban road networks. Unlike dense urban traffic, maritime movements exhibit inherent randomness, heterogeneity, and severe zero inflation (intermittent bursts of activity followed by long periods of inactivity).

The paper identifies that conventional Spatio-Temporal Graph Neural Networks (ST-GNNs), while successful in urban domains, degrade significantly in this maritime context. When trained on highly sparse data, these models tend to converge toward conservative near-zero predictions, failing to capture critical non-zero traffic spikes. While Zero-Inflated Negative Binomial (ZINB) models have been proposed to handle sparsity, the paper notes their limitations: their two-part formulation (separately modeling zero generation and count processes) can yield suboptimal results if zero-generation assumptions are violated, and they often struggle to capture the long-tailed behavior of sudden traffic bursts. Furthermore, existing Tweedie-based approaches in other domains often rely on surrogate training objectives (e.g., lower-bound approximations) to bypass the intractable infinite series in the Tweedie likelihood, introducing complexity and approximation errors.

2. Methodology

The authors propose a model-agnostic learnable Tweedie head designed to be attached as a plug-and-play output module to arbitrary ST-GNN backbones. The core components of the methodology are:

Graph Formulation: The maritime system is modeled as a directed graph $G=(V, E, A)$ , where nodes represent port and anchorage zones, and edges represent navigational connections derived from vessel trajectory transitions.
Tweedie Modeling: Instead of the two-part ZINB mechanism, the model utilizes the Tweedie exponential dispersion family (specifically the compound Poisson-Gamma distribution for $1 < p < 2$ ). This distribution naturally models zero outcomes and continuous positive values without requiring an explicit zero-gating mechanism.
Learnable Variance Power: A key innovation is the prediction of a node-level variance power parameter $p$ . The head splits the backbone features to predict the mean $\mu$ and the power $p$ for each node. This allows the model to adaptively capture spatial heterogeneity, ranging from near-Poisson traffic ( $p \approx 1.1$ ) to overdispersed, Gamma-like traffic ( $p \approx 1.9$ ).
Training Objective (Unit Deviance): To avoid the intractable infinite series required for exact Tweedie log-likelihood, the authors optimize the Tweedie unit deviance. This provides a closed-form objective function that ensures stable backpropagation without relying on lower-bound approximations. The loss function is defined as:
$d(y, \mu, p) = 2 \left( \frac{y^{2-p}}{(1-p)(2-p)} - \frac{y \mu^{1-p}}{1-p} + \frac{\mu^{2-p}}{2-p} \right)$
(with the first term set to 0 when $y=0$ ).
Inference: The model outputs the Tweedie mean $\hat{y} = \mu$ as the point forecast.

3. Experimental Setup

Dataset: Real-world Automatic Identification System (AIS) data from the Port of Los Angeles and Long Beach (LA/LB) from September to November 2024.
Graph Construction: 30 polygonal nodes (terminals and anchorages) were extracted via DBSCAN clustering of stationary AIS signals. 198 directed edges were constructed based on vessel transition counts.
Baselines: The proposed head was tested on three representative ST-GNN backbones: STGCN, DCRNN, and Graph WaveNet (GWNet).
Comparisons: Each backbone was evaluated with three output heads:
1. Base: Standard regression (MAE loss).
2. +ZINB: A ZINB head.
3. +Tweedie (Ours): The proposed learnable Tweedie head.
Metrics: Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) were reported for both "All" data and "Non-zero" data (to mitigate the bias of frequent zeros).

4. Key Results

Performance on Non-Zero Events: Conventional ST-GNNs (Base) achieved the lowest overall MAE but performed poorly on non-zero events, confirming their tendency to under-predict traffic spikes. Both +ZINB and +Tweedie improved Non-zero MAE, demonstrating better sensitivity to actual traffic activity.
RMSE Superiority: The proposed +Tweedie method consistently achieved the lowest RMSE and Non-zero RMSE across all three backbones. Since RMSE penalizes large errors more heavily, this indicates the Tweedie head is superior at capturing the magnitude of sudden traffic surges compared to both the Base and ZINB models.
Qualitative Analysis: Visualizations of 1-step-ahead forecasts showed that while the Base model produced flatlined, conservative predictions, and the ZINB model often underestimated peak magnitudes (due to the discounting effect of the zero-inflation gate $\pi$ ), the Tweedie head produced sharper, more proportional predictions that closely tracked the rise and fall of traffic spikes in both anchorage and terminal zones.

5. Significance and Claims

The paper claims that the proposed model-agnostic learnable Tweedie head offers a robust solution for sparse maritime traffic forecasting by:

Eliminating Approximation Errors: By optimizing the closed-form unit deviance rather than relying on surrogate likelihood approximations, the method ensures stable training without the complexity of infinite-series summations.
Capturing Heterogeneity: The learnable node-level variance power $p$ allows the model to adapt to the specific statistical characteristics of different port areas.
Improving Reliability: The framework significantly improves the prediction of non-zero traffic events, providing a more reliable signal for practical maritime traffic control than conventional ST-GNNs or ZINB-based approaches.

Limitations and Future Work:
The authors modestly note that while the method improves spike prediction, it still tends to remain conservative during the most extreme events. The study is limited to a single region (LA/LB) and a specific three-month period with hourly aggregation. Future work is proposed to include cross-region evaluation, incorporation of exogenous information (e.g., weather, schedules), dynamic graph updates, and the exploration of tail-aware objectives or hybrid approaches to better handle extreme traffic spikes.