LLM-Grounded Explainability for Port Congestion Prediction via Temporal Graph Attention Networks

This paper introduces AIS-TGNN, a framework that combines Temporal Graph Attention Networks with structured Large Language Models to predict port congestion using AIS data while generating operationally interpretable, evidence-grounded natural language explanations that maintain high predictive accuracy and directional consistency.

The Gist

Imagine the world's shipping ports as the heart valves of global trade. When these valves get clogged, the entire body (our supply chain) starts to feel the pain: empty shelves, delayed packages, and rising prices.

For a long time, computers have been good at predicting when these clogs might happen, but they've been terrible at explaining why. It's like a weather app saying, "It will rain tomorrow," but refusing to tell you if it's because of a cold front, a hurricane, or just a leaky roof. Port managers need to know the "why" so they can take action.

This paper introduces a new system called AIS-TGNN that acts like a super-smart detective who not only predicts the traffic jam but also writes a clear, readable report explaining exactly what caused it.

Here is how it works, broken down into simple parts:

1. The Detective's Map: Turning Ships into a Living Graph

Imagine the ocean around the Port of Los Angeles and Long Beach is a giant chessboard.

• The Squares: Instead of chess pieces, each square on the board represents a patch of water.
• The Pieces: The "pieces" are the thousands of ships moving around, sending out digital signals (called AIS) that tell us where they are and how fast they are going.
• The Connections: The system doesn't just look at one ship in isolation. It draws invisible lines between neighboring squares. If a square next door is getting crowded, it "whispers" to the current square, saying, "Hey, things are slowing down over here!"

This creates a Temporal Graph, which is just a fancy way of saying: "A map that changes every day, where every spot knows what its neighbors are doing."

2. The Brain: The "Temporal Graph Attention Network" (TGAT)

Now, imagine a super-brain (the AI model) looking at this map.

• Old Way: Previous models were like a person shouting, "Everyone is loud!" They treated every ship and every square the same, averaging everything out.
• The New Way (TGAT): This model is like a smart traffic cop. It uses "Attention." It asks: "Which specific neighbors are actually causing the trouble?"
  • Maybe the square to the North is full of slow-moving tankers.
  • Maybe the square to the East has a sudden spike in cargo ships.
  • The model learns to pay extra attention to the specific neighbors that matter most for that specific day.

By focusing on the right neighbors, it predicts congestion much better than the old methods.

3. The Translator: The "LLM" (Large Language Model)

This is the magic part. Usually, AI gives you a number (e.g., "85% chance of jam"). But port managers don't speak "numbers"; they speak "stories."

The system takes the AI's internal notes (like "The North neighbor is 90% crowded" and "Ships are moving at 2 knots") and feeds them into a translator bot (a Large Language Model, like the one powering this chat).

• The Rule: The translator is strictly forbidden from making things up. It can only use the facts the AI gave it.
• The Output: Instead of a number, it writes a Risk Report that looks like this:

  "We predict a traffic jam tomorrow. Why? Because the ships in the northern grid are moving very slowly (like a turtle), and the square next door is packed with tankers. If we could speed up those northern ships, the jam might clear."

4. The "Truth Check"

The researchers were worried the translator might start hallucinating (making up fake reasons). So, they built a Truth Check system.

• They asked the AI to write 100 reports.
• They checked every single sentence to see if the "reason" matched the "data."
• The Result: The AI was 99.6% consistent. It almost never lied. It faithfully translated the math into human language.

Why Does This Matter?

Think of it like a medical diagnosis:

• Old System: "You have a 70% chance of being sick." (Scary, but not helpful).
• New System: "You have a 70% chance of being sick because your temperature is high and your white blood cell count is up. Here is what you should do."

For port managers, this means they can stop guessing. If the AI says, "The jam is coming because of the tankers in the North," the manager can send a tugboat to help those specific tankers before the whole port grinds to a halt.

In a Nutshell

This paper built a system that combines math (to predict the future) with language (to explain the present). It turns a complex web of ship movements into a clear, trustworthy story, helping the world keep its supply chains flowing smoothly.

Technical Summary

Here is a detailed technical summary of the paper "LLM-Grounded Explainability for Port Congestion Prediction via Temporal Graph Attention Networks."

1. Problem Statement

Port congestion at major maritime hubs (e.g., Los Angeles/Long Beach) disrupts global supply chains. While existing AI systems can predict congestion, they suffer from a significant "explainability gap."

• Limitation of Current Models: Most state-of-the-art models (e.g., GNNs, LSTMs) prioritize aggregate accuracy metrics (AUC, F1) but fail to provide actionable, node-level insights in natural language.
• Limitation of Post-hoc Explanations: Traditional explainability tools (SHAP, GNNExplainer) output numerical scores or abstract subgraphs that are difficult for non-technical stakeholders (terminal operators, logistics planners) to interpret.
• The Need: There is a critical need for a system that not only predicts congestion escalation with high accuracy but also generates faithful, auditable, and operationally interpretable natural language reports grounded in the model's internal evidence.

2. Methodology: The AIS-TGNN Framework

The authors propose AIS-TGNN, an end-to-end framework that couples a Temporal Graph Attention Network (TGAT) with a structured Large Language Model (LLM) reasoning module.

A. Data and Graph Construction

• Data Source: NOAA AIS broadcasts from the Port of Los Angeles and Long Beach (Jan–June 2023).
• Spatial Discretization: The study area is divided into a $0.1^\circ \times 0.1^\circ$grid (approx. 11km$\times$ 11km).
• Graph Structure:
  • Nodes: Active grid cells containing AIS observations.
  • Edges: Fixed-topology $k$-nearest neighbor graph ($k=8$) based on Euclidean distance.
  • Features ($X_t$): 10 kinematic and traffic-density features per node (e.g., mean speed over ground, slow-vessel ratio, anchor ratio, vessel count, cargo/tanker ratios).
• Label Definition: A binary label ($y=1$) indicates congestion escalation, defined as a cell's "slow-vessel ratio" increasing from day $t$ to $t+1$. The dataset is highly imbalanced (13.5% positive rate).

B. Temporal Graph Attention Network (TGAT) Predictor

• Architecture: The model processes a chronological sequence of daily graph snapshots.
• Mechanism:
  • Uses Multi-head Graph Attention Layers to adaptively weight the influence of neighboring nodes.
  • Temporal Context: Node embeddings from time $t$ are concatenated with raw features at $t+1$ to condition predictions on recent history without using separate recurrent components (like RNNs).
  • Output: A binary classification logit converted to a probability of congestion escalation.
• Training: Trained with a weighted binary cross-entropy loss (weight = 6.74) to handle class imbalance. A strict chronological split (70% train, 15% val, 15% test) prevents temporal data leakage.

C. LLM Explanation Module

• Evidence Extraction: After prediction, the system extracts "evidence records" for each node:
  1. Feature Evidence: Top-5 features ranked by absolute z-score, including their correlation direction (positive/negative) with congestion.
  2. Spatial Evidence: Top-2 spatial neighbors based on learned attention weights, along with their most prominent features.
• Structured Prompting: These records are fed into GPT-4o-mini via a constrained prompt. The LLM is instructed to act as a maritime risk analyst and generate a report strictly based on the provided evidence, prohibiting external hallucinations.
• Output Schema: The LLM generates a JSON-formatted report with six sections:
  1. Target feature drivers.
  2. Neighbor influence.
  3. Risk summary.
  4. Counterfactual suggestions.
  5. Confidence and uncertainty.
  6. Limitations.

D. Validation Protocol

• Directional-Consistency Check: A quantitative metric verifying that the risk direction stated in the LLM's text (e.g., "increases risk") matches the statistical sign of the underlying evidence (z-score sign and point-biserial correlation).

3. Key Contributions

1. AIS-TGNN Dataset & Pipeline: An open-source pipeline transforming raw AIS data into 89 spatiotemporal graph snapshots with ~30,200 labeled node-day samples.
2. High-Performance Predictor: Implementation of a TGAT model achieving superior performance on a highly imbalanced dataset using strict chronological splits.
3. Bridging the Explainability Gap: A novel method to extract internal model evidence (attention weights, z-scores) and inject it as structured constraints into an LLM, ensuring the generated text is mathematically grounded.
4. Auditable Risk Reporting: Demonstration that LLMs can generate multi-section, domain-specific risk reports with 99.6% directional consistency, providing a blueprint for transparent AI in maritime logistics.

4. Experimental Results

The framework was evaluated against Linear Regression (LR) and Graph Convolutional Networks (GCN) baselines.

Model	AUC	AP (Average Precision)	F1 Score	Recall
LR (No Graph)	0.713	0.300	0.375	0.480
GCN (Static Graph)	0.759	0.326	0.383	0.445
AIS-TGNN (Proposed)	0.761	0.344	0.398	0.504

• Performance: TGAT outperformed all baselines. While the AUC gain over GCN was marginal (+0.002), the Recall improved significantly by +13.2% (0.504 vs 0.445).
  • Significance: In port operations, missing a true congestion event (False Negative) is costlier than a false alarm. Higher recall is the critical deployment metric.
• Explainability: The LLM-generated reports achieved a 99.6% directional consistency rate (498/500 judgments), proving the explanations faithfully reflected the model's statistical evidence.
• Feature Analysis: Mean Speed Over Ground (SOG) and Slow Ratio were consistently the top drivers, aligning with physical intuition and dataset-level correlations.

5. Significance and Conclusion

• Operational Deployability: This work moves beyond "black box" AI by providing a practical pathway for Explainable AI (XAI) in maritime risk management. It allows stakeholders to understand why a congestion risk is predicted, not just that it is predicted.
• Trust and Auditability: By grounding LLM generation in verifiable model internals (attention weights and z-scores), the system mitigates the risk of hallucination and ensures the reports are auditable.
• Future Directions: The authors suggest integrating exogenous data (weather, tides, berth schedules) to further improve recall and extending the framework to multi-step-ahead predictions.

In summary, AIS-TGNN successfully demonstrates that high-accuracy spatiotemporal prediction and human-interpretable, faithful natural language explanations can be achieved simultaneously, offering a robust tool for proactive supply chain risk management.