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

Spatio-Temporal Graphs Beyond Grids: Benchmark for Maritime Anomaly Detection Jeehong Kim∗ Graduate School of Data Science Seoul National University williamkim10@snu.ac.kr Youngseok Hwang∗ Graduate School of Data Science Seoul National University yshwang35@snu.ac.kr Minchan Kim Graduate School of Data Science Seoul National University mmm5373@snu.ac.kr Sungho Bae Graduate School of Data Science Seoul National University sunghobae@snu.ac.kr Hyunwoo Park Graduate School of Data Science Seoul National University hyunwoopark@snu.ac.kr Abstract Spatio-temporal graph neural networks (ST-GNNs) have achieved notable success in structured domains such as road traffic and public transportation, where spatial entities can be naturally represented as fixed nodes. In contrast, many real-world systems including maritime traffic lack such fixed anchors, making the construction of spatio-temporal graphs a fundamental challenge. Anomaly detection in these non-grid environments is particularly difficult due to the absence of canonical reference points, the sparsity and irregularity of trajectories, and the fact that anomalies may manifest at multiple granularities. In this work, we introduce a novel benchmark dataset for anomaly detection in the maritime domain, extending the Open Maritime Traffic Analysis Dataset (OMTAD) into a benchmark tailored for graph-based anomaly detection. Our dataset enables systematic evaluation across three different granularities: node-level, edge-level, and graph-level anomalies. We plan to employ two specialized LLM-based agents: Trajectory Synthesizer and Anomaly Injector to construct richer interaction contexts and generate semantically meaningful anomalies. We expect this benchmark to promote reproducibility and to foster methodological advances in anomaly detection for non-grid spatio-temporal systems. 1 Introduction Spatio-temporal graph neural networks (ST-GNNs) have been extensively studied in domains such as road traffic forecasting and public transportation systems [19, 3, 2]. A common characteristic of these applications is that the underlying spatial entities like road intersections, bus stops, or subway stations can be naturally defined as fixed nodes. This inherent grid-like structure makes the construction of spatio-temporal graphs straightforward and facilitates the modeling of both spatial dependencies and ∗Equal contribution (co-first authors). 39th Conference on Neural Information Processing Systems (NeurIPS 2025) Workshop: AI for Science: The Reach and Limits of AI for Scientific Discovery. arXiv:2512.20086v1 [cs.LG] 23 Dec 2025 temporal dynamics. Consequently, anomaly detection in such structured environments has received significant attention and demonstrated promising results [1]. However, there are many cases both in real-world and scientific domains where situations do not conform to these assumptions. In particular, there exist domains where fixed spatial anchors are absent or physically ambiguous. The maritime environment represents one of the most prominent examples: unlike road traffic systems, the open sea does not provide natural fixed nodes such as intersections or road segments. Although artificial proxies such as waypoints, port coordinates, or grid discretizations can be imposed, these methods are often ad hoc and fail to capture the continuous and dynamic nature of vessel trajectories. This fundamental challenge renders the construction of a meaningful spatio-temporal graph a non-trivial task. We expect that such non-grid spatio-temporal systems will become increasingly common, not only in maritime monitoring but also in emerging domains such as drone swarms and aerial traffic management. Performing anomaly detection in these settings is even more challenging. First, the lack of fixed spatial anchors complicates the definition of normal versus abnormal interactions among moving entities. Second, the inherent sparsity and irregularity of the trajectories make it difficult to design robust models. Third, anomalous patterns may manifest at multiple levels: individual entities (node- level anomalies), unusual pairwise interactions (edge-level anomalies), or entire subgroups behaving abnormally (graph-level anomalies). These challenges highlight the need for systematic benchmarks that enable rigorous evaluation and foster methodological innovations [7]. There are several Marine datasets To address this gap, in this paper we introduce a novel benchmark dataset for anomaly detection in the maritime domain. Our dataset is designed to support anomaly detection tasks at three granularities: (i) node-level anomalies, capturing abnormal single-entity behaviors, (ii) edge-level anomalies, reflecting irregular inter-entity interactions, and (iii) graph-level anomalies, identifying collective abnormal events. Inspired by recent advances in graph anomaly detection across node-, edge-, and graph-level settings [5], we aim to provide a unified testbed that allows the communi

This content is AI-processed based on open access ArXiv data.