Skip to content

Spatio-Temporal Graphs Beyond Grids: Benchmark for Maritime Anomaly Detection

🕒 Published (v1): 2025-12-23 06:28 UTC · Source: Arxiv · Venue: NEURIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This workshop paper introduces a benchmark dataset for graph-based anomaly detection in maritime traffic by extending the Open Maritime Traffic Analysis Dataset (OMTAD). Two LLM-based agents—a Trajectory Synthesizer and an Anomaly Injector—are used to populate sparse vessel graphs and inject semantically meaningful anomalies at node, edge, and graph granularities.

Problem

Existing ST-GNN benchmarks assume fixed spatial anchors (road intersections, transit stops), but maritime traffic operates on an open sea with no natural fixed nodes. OMTAD provides only normal vessel trajectories with no anomaly labels, and many trajectories are isolated with no nearby neighbors—preventing graph construction and systematic evaluation of multi-level anomaly detection.

Method

The authors extend OMTAD via a two-agent LLM pipeline under a Coordinator orchestrator: 1. Trajectory Synthesizer: For each focal vessel, generates synthetic "virtual neighbor" trajectories by perturbing SOG, COG, and geolocation within realistic bounds, ensuring graph connectivity even in sparse regions. If real neighbors exist, they are used directly. 2. Anomaly Injector: Takes high-level natural language prompts (e.g., "unusual speed changes," "risky encounters," "group loitering"), parses them into structured intent, and modifies the spatio-temporal graph to inject node-level (single vessel kinematics), edge-level (inter-vessel interaction), or graph-level (collective group behavior) anomalies with attached labels and rationale text.

Both agents share a common perception schema covering AIS fields (MMSI, lat/lon, SOG, COG), derived kinematic rates (ΔSOG/Δt, ΔCOG/Δt), environmental bins, and provenance IDs.

Preliminary experiments use OPTICS clustering to form dynamic vessel graphs, comparing LSTM/Transformer baselines against GNN-augmented variants under varying trajectory anomaly ratio (r_traj ∈ {0.1, 0.5}) with fixed node anomaly ratio r_node = 0.5, with kinematic anomalies injected via >3σ perturbations of SOG/COG rates.

Key Contributions

  • A multi-granularity benchmark extending OMTAD for node-, edge-, and graph-level maritime anomaly detection.
  • A two-agent LLM pipeline (Trajectory Synthesizer + Anomaly Injector) for principled, prompt-driven dataset augmentation and anomaly injection.
  • A shared structured perception schema (AIS + derived kinematics + environmental context) enabling reproducible, deterministic pipeline execution.
  • Preliminary empirical validation showing GNN-integrated models consistently outperform purely temporal baselines on the injected anomaly task.

Results

  • GNN-integrated models outperform pure temporal baselines (LSTM, Transformer) consistently across both trajectory anomaly ratios (r_traj = 0.1 and 0.5).
  • Results are shown for LSTM+GNN vs. LSTM and Transformer+GNN vs. Transformer under fixed r_node = 0.5; specific metric values are presented in figures (Fig. 1a, 1b) but numerical scores are not reported in the text.
  • No comparison against prior maritime anomaly detection methods (GeoTrackNet, TrAISformer, graph attention networks) is included at this stage.

Limitations

  • Benchmark is not yet finalized or publicly released; the full pipeline and dataset remain works-in-progress as of publication.
  • Preliminary experiments cover only graph-level anomaly detection with kinematic (SOG/COG) perturbations—far simpler than real-world maritime anomalies (illegal rendezvous, AIS spoofing, loitering).
  • No quantitative results with named metrics (AUC, F1, etc.) are reported; findings are presented only as figures without tabulated numbers.
  • Synthetic neighbor generation and prompt-driven anomaly injection introduce distributional artifacts not validated against ground-truth anomaly events.
  • LLM agent configurations are detailed in an appendix but the specific models used are not disclosed.
  • Node- and edge-level detection tasks are described but not yet evaluated.

Relevance to Agentic AI / LLM Agents

This paper is a direct application of multi-agent LLM orchestration to scientific dataset construction: two specialized agents (Synthesizer, Injector) operate under a Coordinator, decomposing a complex data curation task into distinct roles with shared structured inputs. It exemplifies the "LLM-as-data-generator" paradigm where agents replace brittle rule-based procedures with flexible prompt-driven generation, a pattern increasingly relevant for building evaluation benchmarks in domains with scarce labeled data. The agent design—structured perception schema, provenance tracking, rationale-annotated outputs—demonstrates engineering patterns applicable to agentic pipelines in scientific domains beyond maritime. Practitioners tracking agentic AI should note the architectural parallel to tool-using agent systems with a central Coordinator dispatching subtasks to specialized modules.