Graph Enhanced Trajectory Anomaly Detection

• Graph Enhanced Trajectory Anomaly Detection (GETAD) is a framework that models trajectories as sequences of map-matched road segments within a directed road network.
• It integrates graph attention networks, Transformer decoders, and multiobjective loss combining autoregressive and link prediction to capture local structure and historical travel patterns.
• GETAD employs Confidence Weighted Negative Log Likelihood for anomaly scoring, demonstrating enhanced detection of subtle abnormal road movements in real-world datasets.

Searching arXiv for the GETAD paper and closely related graph-based anomaly detection work. arXiv search: "Graph Enhanced Trajectory Anomaly Detection" Graph Enhanced Trajectory Anomaly Detection (GETAD) is a framework for trajectory anomaly detection that models trajectories not merely as sequences of sampled locations or staypoints, but as sequences of map-matched road segments embedded in a directed road network. In this formulation, the movement space is represented explicitly by road network topology, road segment semantics, and historical travel patterns. GETAD combines a Graph Attention Network for road-aware embeddings, graph-based positional encodings that reflect network layout, a Transformer-based decoder for sequential movement modeling, a multiobjective training loss combining autoregressive prediction and supervised link prediction, and Confidence Weighted Negative Log Likelihood (CW-NLL) for anomaly scoring. The framework is designed to improve detection of subtle anomalies in road-constrained environments, where Euclidean trajectory analysis can neglect network connectivity and movement constraints (Mbuya et al., 22 Sep 2025).

1. Problem setting and conceptual scope

Trajectory anomaly detection concerns the identification of unusual and unexpected movement patterns. Within the formulation adopted by GETAD, the central criticism of earlier approaches is that they only consider limited aspects of trajectory data by treating trajectories as sequences of sampled locations, with sampling determined by positioning technology such as GPS, or by high-level abstractions such as staypoints. In addition, those approaches analyze trajectories in Euclidean space and thereby neglect the constraints and connectivity information of the underlying movement network, such as road or transit networks (Mbuya et al., 22 Sep 2025).

GETAD is therefore explicitly network-aware. Its operative assumption is that anomalous movement is not adequately characterized by geometric deviation alone; it must also be evaluated relative to feasible transitions, segment semantics, and historically observed travel behavior. This places GETAD within a broader line of research that uses graph structure to model dependencies that conventional sequence-only methods underrepresent. Earlier work on graph time series modeled temporal correlation with recurrent neural networks and variational inference while using graph convolutional networks to capture spatial information, demonstrating that graph structure can materially improve anomaly detection in traffic flow data (Hsu, 2017). Likewise, anomaly detection in multi-agent trajectories for automated driving used a spatio-temporal graph auto-encoder to jointly learn multiple trajectories of a dynamic number of agents and then estimate a density function in latent space, showing that interaction structure can be critical for detecting abnormal manoeuvres (Wiederer et al., 2021).

A common misconception is that graph enhancement in trajectory anomaly detection is equivalent to adding an adjacency matrix to an otherwise unchanged sequence model. GETAD is structurally richer than that interpretation. Its graph enhancement includes a directed road-segment graph, node features capturing both physical and behavioral statistics, graph attention with transition probabilities, graph-based positional encoding derived from shortest-path distances, and an auxiliary link prediction objective that preserves network connectivity in the learned representation (Mbuya et al., 22 Sep 2025).

2. Road-network representation and road-aware embeddings

GETAD represents the road network as a directed graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$, where each node $v \in \mathcal{V}$ is a directed road segment and each edge $(v_i, v_j) \in \mathcal{E}$ denotes a feasible transition from segment $v_i$ to $v_j$. Each segment is associated with a feature vector containing physical attributes and historical or usage statistics. The physical attributes include length, type, speed limit, in-degree, out-degree, and number of lanes. The historical or behavioral statistics include normalized visit frequency (Mbuya et al., 22 Sep 2025).

The road-aware embedding module is implemented with a Graph Attention Network. For each segment, the initial feature vector is denoted $\mathbf{h}^0_i \in \mathbb{R}^{d_0}$ and is formed by concatenating physical and behavioral statistics. Directed transitions are quantified by an empirical transition probability,

$$p_{v_i v_j} = \frac{\text{count}(v_i \rightarrow v_j)}{\text{count}(v_i)}.$$

This quantity is incorporated directly into the attention mechanism. For each attention head $k$ at layer $l$, GETAD computes

$$e_{v_i v_j}^{l,k} = (\mathbf{a}_1^{l,k})^\top\, \text{LeakyReLU} (\mathbf{W}_1^{l,k} \mathbf{h}_i^l + \mathbf{W}_2^{l,k} \mathbf{h}_j^l + p_{v_i v_j} \cdot \mathbf{a}_2^{l,k}),$$

followed by attention normalization and neighbor aggregation,

$v \in \mathcal{V}$0

The head outputs are concatenated and passed through a nonlinearity: $v \in \mathcal{V}$1 According to the framework description, this allows segment embeddings to encode local structure, including connections, degrees, and immediate neighbors, together with global semantics such as type, usage patterns, and transition frequencies (Mbuya et al., 22 Sep 2025).

This design aligns with a wider methodological trend in graph-based anomaly detection: relational structure is not treated as a static scaffold but as a component of probabilistic or neural representation learning. In "Graph-Augmented Normalizing Flows for Anomaly Detection of Multiple Time Series" (Dai et al., 2022), for example, a Bayesian network factorizes the joint probability of multiple series into conditional densities modeled by normalizing flows, illustrating a different but related use of graph structure to encode interdependence.

3. Graph-based positional encoding and sequential decoding

A distinctive component of GETAD is graph-based positional encoding (GPE). Standard positional encodings rely on token index order, but GETAD instead encodes spatial and topological relations in the underlying road network. For a trajectory $v \in \mathcal{V}$2, the method computes a shortest-path distance matrix $v \in \mathcal{V}$3 of size $v \in \mathcal{V}$4, where $v \in \mathcal{V}$5 is the minimal hop distance between $v \in \mathcal{V}$6 and $v \in \mathcal{V}$7 in $v \in \mathcal{V}$8. Each distance value is mapped through a learnable lookup table,

$v \in \mathcal{V}$9

and the positional encoding for token $(v_i, v_j) \in \mathcal{E}$0 is

$(v_i, v_j) \in \mathcal{E}$1

The stated purpose is to encode network proximity and topological closeness between positions rather than mere order in the sequence (Mbuya et al., 22 Sep 2025).

The sequential model is a Transformer decoder. At each position, the decoder input is the sum of the GAT-derived segment embedding and the graph-based positional encoding,

$(v_i, v_j) \in \mathcal{E}$2

Autoregressive masking ensures that each position attends only to previous or current positions, preserving causal structure. The decoder then predicts a probability distribution over the next possible road segment (Mbuya et al., 22 Sep 2025).

This combination of graph structure and sequence modeling places GETAD in a family of hybrid spatiotemporal architectures, but its emphasis is specifically road-constrained movement. Related work in human-centric video anomaly detection, such as TSGAD, also couples graph-based context reasoning with sequence prediction and a two-stream design based on reconstruction and forecasting. That work uses dynamic graphs and trajectory prediction in a pose-based video anomaly detection setting and reports comparable results with state-of-the-art methods, which suggests a broader relevance of graph-informed sequence models for motion anomalies beyond road networks (Noghre et al., 2024).

A plausible implication is that GETAD’s graph-based positional encoding addresses a representational gap that standard Transformer positional encodings leave unresolved in network-constrained mobility data. The framework description does not frame this as a theoretical guarantee, but it does identify the encoding as a mechanism for capturing true spatial and topological relationships rather than sequence index alone (Mbuya et al., 22 Sep 2025).

4. Multiobjective training and anomaly scoring

GETAD is trained with a multiobjective loss that combines autoregressive next-segment prediction and supervised link prediction. The cross-entropy term for sequential prediction is

$(v_i, v_j) \in \mathcal{E}$3

The link prediction term is

$(v_i, v_j) \in \mathcal{E}$4

where $(v_i, v_j) \in \mathcal{E}$5 is the dot-product similarity between segment embeddings and $(v_i, v_j) \in \mathcal{E}$6 indicates whether transition $(v_i, v_j) \in \mathcal{E}$7 exists in the training data. The total objective is

$(v_i, v_j) \in \mathcal{E}$8

This training strategy is described as encouraging representations that both fit sequential movement data and preserve empirically observed and plausible transitions in the topology (Mbuya et al., 22 Sep 2025).

For anomaly scoring, GETAD introduces Confidence Weighted Negative Log Likelihood (CW-NLL). Let $(v_i, v_j) \in \mathcal{E}$9 denote the negative log-likelihood for the true next segment at token $v_i$0, and let $v_i$1 denote the model’s softmax distribution over next segments. Its entropy is

$v_i$2

with maximum entropy $v_i$3, where $v_i$4 is the number of segments. Confidence is defined as

$v_i$5

and the final trajectory score is

$v_i$6

The stated rationale is that raw negative log-likelihood or perplexity can be misleading when the model is uncertain because several continuations are plausible. CW-NLL therefore emphasizes unlikely events that occur under high-confidence predictions and discounts ambiguous low-confidence cases (Mbuya et al., 22 Sep 2025).

This scoring strategy differs from other anomaly criteria in the literature. ADAMM, for instance, models daily urban mobility trajectories as node- and edge-attributed directed multi-graphs, fuses graph and metadata embeddings, and uses a cluster-based anomaly objective in a multi-centroid latent space,

$v_i$7

arguing that trajectory data is inherently multi-modal (Sotiropoulos et al., 2023). GETAD instead keeps the anomaly decision tied to sequential likelihood but modifies that likelihood by prediction confidence, thereby prioritizing structurally coherent deviations in the learned road-aware sequence model (Mbuya et al., 22 Sep 2025).

5. Empirical evaluation and observed behavior

The evaluation reported for GETAD uses both a synthetic dataset and a real-world dataset. The synthetic setting is an agent-based “Pattern of Life” simulation using San Francisco’s road network, and the real-world dataset is the Porto Taxi dataset. The experiments consider constrained anomalies, characterized as subtle, small detours, and unconstrained anomalies, characterized as larger, more obvious deviations (Mbuya et al., 22 Sep 2025).

Across these settings, GETAD is reported to outperform the baselines SAE, VSAE, GM-VSAE, and LM-TAD in both F1 and PR-AUC for both subtle and large detour settings. The framework is described as being especially effective for subtle or constrained anomalies, where deviations are minimal and difficult to detect. The anomaly score CW-NLL is also reported to consistently outperform raw NLL and perplexity, particularly by improving precision through discounting low-confidence, ambiguous cases (Mbuya et al., 22 Sep 2025).

On the Porto dataset with constrained anomalies, the reported table gives the following values:

Method	F1	PR-AUC
SAE	.575	.566
VSAE	.579	.564
GM-VSAE	.677	.664
LM-TAD	.695	.732
GETAD	.701	.756

The ablation findings identify both graph-based positional encoding and structural link supervision as crucial for detecting subtle, near-normal detours. The description states that relying only on sequence modeling or on basic positional encodings significantly reduces the ability to discern contextually abnormal routes (Mbuya et al., 22 Sep 2025).

This reported behavior is consistent with related trajectory anomaly literature in which interaction structure or richer graph structure improves anomaly sensitivity. In multi-agent driving, the spatio-temporal graph auto-encoder with kernel density estimation achieved the highest AUROC of 86.28% and the highest AUPR scores among the methods compared on the MAAD dataset, with the authors explicitly attributing gains to modeling interactions among agents (Wiederer et al., 2021). Although the task setting differs from road-segment sequence modeling, both results emphasize that anomalies can be subtle relative to local geometry while still being abnormal relative to a structured movement environment.

6. Relation to adjacent research traditions

GETAD occupies an intersection between trajectory anomaly detection, graph representation learning, and multivariate time-series anomaly detection. One adjacent line of work models trajectories or mobility records as attributed graphs rather than as simple sequences. ADAMM handles directed graphs with multi-edges and self-loops, fuses metadata and graph-level representation learning in a unified end-to-end architecture, and validates the approach on general-ledger journal entries and human GPS trajectories from thousands of individuals. In the urban mobility setting, each agent-day is modeled as a directed multi-graph whose nodes are visited locations and whose directed edges or multi-edges encode trips with attributes such as duration and distance. On the MobiNet dataset, ADAMM reports an average AUROC of 0.787 and an average AUPRC of 0.443, compared with the best baseline values of 0.730 and 0.208, respectively (Sotiropoulos et al., 2023).

A second adjacent tradition concerns graph-enhanced multivariate time-series anomaly detection. PMGC uses a long-term static graph together with a series of short-term instance-wise dynamic graphs, regulated through a graph cohesion loss, and introduces a prospective graphing strategy to incorporate current time-series values into graph construction. On five real-world benchmarks, PMGC is reported to significantly outperform both non-graph TSAD baselines and single-graph GNN TSAD baselines, with average improvements of about 13% in pointwise F1, 18% in F1-composite, 8% in VUS-PR, and 21% in VUS-ROC (Chen et al., 21 Sep 2025). While PMGC is not a road-network trajectory framework, it provides a closely related example of how multiple graph views can improve anomaly detection when inter-series relations change over time.

A third surrounding tradition is general graph anomaly detection, where the graph itself is the primary object. DDGAD frames anomalies as instability in node representation trajectories under coupled diffusion and reliability-aware consensus dynamics, introducing anomaly signals such as neighbor inconsistency, reliability weight, and dynamical conflict energy (Yang et al., 26 May 2026). CurvGAD introduces curvature-based geometric anomalies through a mixed-curvature graph autoencoder and reports up to 6.5% improvement over state-of-the-art GAD methods on 10 real-world datasets (Grover et al., 12 Feb 2025). These works do not solve road-trajectory anomaly detection directly, but they make clear that graph enhancement can target distinct sources of abnormality: contamination propagation, geometric irregularity, semantic inconsistency, or movement-network incoherence.

Within this landscape, GETAD is specifically defined by road-constrained sequential movement, graph-based positional encoding, and confidence-weighted likelihood scoring (Mbuya et al., 22 Sep 2025).

7. Interpretation, limitations, and research directions

The principal interpretive contribution of GETAD is its insistence that a trajectory should be judged relative to the movement network in which it occurs. In that sense, the framework is not simply a better sequence model; it is a redefinition of the anomaly detection problem for road-constrained mobility. The reported gains on constrained anomalies support the claim that subtle detours can be contextually abnormal even when they are not geometrically extreme (Mbuya et al., 22 Sep 2025).

A common misunderstanding is to equate road awareness with hard feasibility checking. GETAD goes beyond feasible-transition filtering. Its representation includes road segment semantics and historical travel patterns, its decoder models sequential movement autoregressively, and its link prediction loss preserves connectivity information in embedding space. The anomaly signal is therefore not limited to impossible transitions; it also targets low-probability but structurally coherent deviations under a learned model of normal travel (Mbuya et al., 22 Sep 2025).

At the same time, the framework description leaves some broader questions open. The abstract and details emphasize real-world and synthetic evaluation, consistent improvements over existing methods, and enhanced precision for subtle anomalies, but they do not provide a claim of universal superiority across all trajectory domains. This suggests that GETAD is most directly applicable where trajectories are map-matched to a known movement network and where road segment semantics and transition behavior are informative. By contrast, anomaly settings centered on unconstrained Euclidean motion, multi-agent interactions without fixed road topology, or heterogeneous metadata-rich mobility graphs may favor other formulations such as spatio-temporal graph auto-encoders, graph-metadata fusion, or two-stream prediction-and-reconstruction models (Wiederer et al., 2021, Sotiropoulos et al., 2023, Noghre et al., 2024).

A plausible implication is that future extensions of GETAD could draw on themes already visible in adjacent work: richer metadata fusion as in ADAMM, dynamic graph adaptation as in PMGC, or probabilistic density modeling of graph-conditioned trajectories as in GANF. Those implications are not explicit claims of the GETAD framework itself, but they follow naturally from the current research landscape in graph-enhanced anomaly detection (Sotiropoulos et al., 2023, Chen et al., 21 Sep 2025, Dai et al., 2022).