PaperlayerView exact revision on arXiv

Source-linked AI summary

One Pretrained Decision Model Is Enough for Training-Free Generalist Graph Anomaly Detection

Xinyu Zhao, Qingyun Sun, Ziwei Zhang, Tianchen Zhu, Daqing Li, Jianxin Li

arXiv:2610.30777v1cs.LG

TL;DR

Graph anomaly detection faces heterogeneous evidence and cross-graph distribution differences, while training-free approaches often rely on fixed numerical scoring rules. Jev-GAD combines graph-relative evidence with semantically grounded pretrained decision readout and statistical scores. Across seven benchmarks, it achieves average AUROC of 89.18% and average AUPRC of 52.40%.

  • Problem

    Heterogeneous anomaly evidence and distributional differences across graphs challenge generalizable detection, while training-free methods leave the semantic interpretation of evidence largely unexplored.

    Source: §1 · pp. 1–2

  • Method

    Jev-GAD combines graph-relative structural, attribute, and neighborhood evidence with semantically grounded, non-generative decision readout from one frozen pretrained model and equal-weight rank fusion with statistical scores.

    Source: §1 · p. 2

  • Results

    89.18% average AUROC and 52.40% average AUPRC across seven benchmarks; Jev-GAD surpasses strongest-baseline macro-averages by 7.09 and 15.44 percentage points, respectively.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    The results support pretrained decision readout and semantic evidence grounding as effective components for training-free graph anomaly detection across the evaluated benchmarks.

    Source: §1 · p. 2, §6 · p. 9

    The setting applies to unseen target graphs without anomaly labels or GAD-specific gradient-based training, with the pretrained decision model kept frozen.

    Source: §3 · p. 3

Abstract

from arXiv · show

Graph anomaly detection (GAD) aims to identify irregular entities in graph-structured data, yet heterogeneous anomaly patterns and distributional differences across graphs challenge generalizable detection. Existing training-free approaches avoid task-specific optimization but primarily rely on predefined numerical scoring rules, leaving the semantic interpretation of heterogeneous anomaly evidence largely unexplored. In this paper, we propose Jev-GAD, a training-free generalist graph anomaly detection framework that bridges heterogeneous numerical evidence and pretrained decision mechanisms. To address heterogeneous anomaly patterns, we construct complementary structural, attribute, and neighborhood evidence and transform heterogeneous statistics into graph-relative anomaly percentiles. To move beyond predefined numerical scoring rules, we leverage a single frozen pretrained language model for evidence-conditioned anomaly assessment. To bridge the semantic gap between numerical evidence and pretrained models, we design a semantic Graph-to-State (G2S) transformation that organizes graph context and semantically labeled node evidence into structured decision states for non-generative option-logit readout. The resulting decision scores are integrated with four complementary statistical scores through equal-weight rank-based fusion, without GAD-specific gradient-based optimization. Extensive experiments on seven graph anomaly detection benchmarks spanning citation, social, and co-review networks demonstrate the effectiveness of Jev-GAD, achieving average AUROC and AUPRC scores of 89.18\% and 52.40\%, respectively. Further analyses validate the contributions of pretrained decision readout and semantic evidence grounding.

1 Introduction

Graph anomalies vary across structural, attribute, and neighborhood evidence, while graph distributions differ across domains; training-free methods avoid task-specific training but rely on fixed numerical scoring. Jev-GAD grounds heterogeneous evidence semantically for pretrained decision readout and combines it with statistical scores.

1 Introduction

  • Training-free approaches avoid GAD-specific parameter training, but existing methods commonly map heterogeneous evidence to predefined numerical scores without conditioning on evidence semantics.Structural rarity and neighborhood inconsistency can have different implications depending on graph context.Source: §1 · pp. 1–2
  • Graph anomalies include unusual structural positions, atypical attributes, and inconsistencies with neighbors, while substantial distributional differences across Web graphs challenge generalization.The motivation includes suspicious entities and irregular interactions in social networks, e-commerce, and recommendation services.Source: §1 · p. 1
  • Graph-relative numerical values alone do not identify whether evidence represents structural rarity, attribute deviation, or neighborhood inconsistency, motivating explicit semantic labels.The paper proposes associating graph-relative evidence with anomaly semantics so frozen pretrained models can assess evidence without graph-specific parameter learning.Source: §1 · p. 2
  • Jev-GAD combines structural, attribute, and neighborhood evidence, graph-relative anomaly percentiles, and a single frozen pretrained model for evidence-conditioned assessment.Its semantic Graph-to-State transformation organizes graph context and semantically labeled node evidence into structured states for non-generative option-logit readout; four statistical scores join decision scores through equal-weight rank-based fusion.Source: §1 · p. 2
  • 89.18% average AUROC and 52.40% average AUPRC across seven benchmarks; Jev-GAD surpasses strongest-baseline macro-averages by 7.09 and 15.44 percentage points, respectively.Further analyses validate pretrained decision readout, semantic evidence grounding, and complementary statistical scoring.Source: §1 · p. 2

2 Related Work

Graph anomaly detection methods capture evidence at different levels, while transferable methods still often require task-specific training or adaptation. Pretrained decision models offer non-generative option-level scoring, but their use with heterogeneous graph evidence remains underexplored.

2.1 Graph Anomaly Detection

  • Traditional detectors use residual modeling, reconstruction, contrastive learning, or community patterns; training-free TFGAD and FreeGAD instead rely on intrinsic graph statistics and predefined scoring.The approaches capture node attributes, neighborhood relations, and higher-order community structure, while benchmarking reveals varied detector characteristics and performance across attributed networks.Source: §2.1 · p. 2

2.2 Pretrained Decision Models

  • Non-generative decision readout obtains option-level scores conditioned on predefined questions and contextual states, unlike free-form generative judging.Option-level scores can be integrated directly with numerical anomaly scorers; applying pretrained decision mechanisms to heterogeneous graph evidence remains relatively underexplored.Source: §2.2 · pp. 2–3

3 Problem Formulation

The paper formulates node-level anomaly detection on attributed graphs as assigning higher scores to anomalous nodes. Its training-free generalist setting applies a frozen pretrained decision model directly to unseen target graphs without anomaly labels or GAD-specific gradient-based training.

3 Problem Formulation

  • The objective is a node-level scoring function that assigns higher scores to anomalous nodes, applied directly to unseen target graphs without anomaly labels or GAD-specific gradient-based training.The pretrained decision model remains frozen, and target anomaly labels are used only for evaluation.Source: §3 · p. 3

4 Methodology

Jev-GAD builds graph-relative structural, attribute, and neighborhood evidence, organizes it into semantically grounded states for pretrained decision readout, and fuses those decisions with complementary statistical scores.

4.1 Heterogeneous Graph Evidence Construction

  • Graph-relative percentiles place structural, attribute, and neighborhood statistics on a common scale of atypicality across target-graph nodes.For potentially anomalous upper- and lower-tail statistics, deviations are measured from the graph-wide median; distance-based deviations are retained directly.Source: §4.1 · p. 3

4.2 Semantic Graph-to-State Representation

  • Graph-to-State (G2S) pairs semantically labeled node evidence with graph-level context, preserving what each value represents for pretrained assessment.The serialized state groups node evidence by anomaly characteristic and includes graph properties such as size, connectivity, and feature characteristics.Source: §4.2 · pp. 3–4

4.3 Pretrained Decision Readout

  • A frozen pretrained model reads anomalous-versus-normal option logits conditioned on each node’s state, producing a decision score in one forward pass without generation or parameter updates.The decision criterion and candidate options remain fixed across nodes and graphs, while the evidence state varies by node.Source: §4.3 · p. 4

4.4 Statistical Scoring and Fusion

  • Four statistical scorers and the pretrained decision score are converted to target-graph percentile ranks and fused with equal weight, 1/5 each.The statistical mechanisms are fitted on unlabeled target-graph evidence, without anomaly-specific supervision or label-dependent weight estimation.Source: §4.4 · pp. 4–5

5 Experiments

The experiments evaluate Jev-GAD across seven attributed graph benchmarks and examine overall effectiveness, component contributions, semantic grounding, evidence choices, robustness, and computational costs.

5.1 Experimental Settings

  • The evaluation covers seven attributed benchmarks, reports AUROC and AUPRC, and compares 15 baselines using unlabeled target evidence and frozen Qwen3-4B decision readout.Jev-GAD requires neither target anomaly labels nor GAD-specific gradient-based optimization at inference; readout uses two fixed candidate decision tokens without autoregressive generation or parameter updates. The statistical branch uses Fixed, ECOD-style, Isolation Forest, and RBF-based OCSVM scores, rank-normalized before equal-weight fusion.Source: §5.1 · pp. 5–6

5.2 RQ1: Overall Effectiveness

  • 89.18% AUROC and 52.40% AUPRC surpass the strongest baseline macro-averages by 7.09 and 15.44 percentage points, respectively.Jev-GAD also achieves the highest AUROC on six of seven graphs and highest AUPRC on five.Source: §5.2 · p. 6, Table 2 · p. 6

5.3 RQ2: Component Ablation

  • Jev-GAD outperforms Decision-only and Statistical-only on all four datasets; its four-graph macro-AUPRC is 59.72% versus 51.41% for Statistical-only and lower numerical alternatives.The numerical alternatives are Mean-7D (49.98%), Max-7D (50.04%), and LOF-7D (42.14%).Source: §5.3 · p. 7, Figure 3 · p. 7

5.4 RQ3: Semantic Evidence Grounding

  • On CiteSeer, Decision-only AUPRC falls from 74.04% with original descriptions to 19.79% with Anonymous Fields and 41.90% with Mismatched Semantics.On Amazon, full Jev-GAD scores 41.26% AUPRC with the original state versus 25.76% and 27.95% under those perturbations, while numerical evidence and statistical scores stay unchanged.Source: Figure 3 · p. 7, §5.4 · p. 8

5.5 RQ4: Evidence and Scoring Analysis

  • Both raw and aligned views yield the highest AUPRC on three of four graphs, while the original statistical backend beats each tested replacement on all four.On Amazon, raw, aligned, and combined views score 22.70%, 37.23%, and 41.26% AUPRC; on CiteSeer, the original backend scores 88.02% versus 51.01%, 54.01%, and 71.39% for alternatives.Source: Figure 3 · p. 7, §5.5 · p. 8

5.6 RQ5: Decision Robustness

  • Across tested prompts, CiteSeer AUPRC ranges from 85.01% to 88.20%; equal-weight λ=0.2 gives highest macro-AUROC, while the endpoints lose 8.31 and 18.90 macro-AUPRC points.Peak AUPRC occurs at λ=0.2 on Amazon and BlogCatalog, 0.3 on Cora, and 0.5 on CiteSeer.Source: §5.6 · pp. 8–9, Figure 4 · p. 8, Figure 5 · p. 8

5.7 RQ6: Computational Efficiency

  • On Amazon, Jev-GAD takes approximately 116 seconds, including 31 seconds of CPU processing and 85 seconds of model inference; on Cora it takes approximately 31 seconds.Frozen Qwen3-4B readout processes approximately 120 nodes per second with about 9.6 GiB peak allocated GPU memory.Source: Figure 6 · p. 8, §5.7 · p. 9

6 Conclusion

Jev-GAD combines semantically grounded pretrained decision readout with complementary statistical scoring for training-free graph anomaly detection. Across seven benchmarks, it achieves average AUROC and AUPRC of 89.18% and 52.40%.

6 Conclusion

  • Jev-GAD organizes graph-relative evidence into semantically grounded inputs for frozen-model decision readout, then rank-fuses those decisions with complementary statistical scores.Across seven benchmarks, it achieves average AUROC and AUPRC of 89.18% and 52.40%; component ablations and robustness analyses validate pretrained decision readout and semantic grounding.Source: §6 · p. 9

A Additional Method Details

Jev-GAD converts graph information into graph-relative evidence, presents selected evidence in semantically structured states for a frozen model, and combines its decision scores with statistical scores through rank fusion.

A.1 Graph Evidence Construction

  • Graph statistics are converted into directionally consistent graph-relative percentiles, with 16 evidence columns spanning structural, attribute, and neighborhood measurements.The aligned feature representation uses a fixed random seed and does not optimize anomaly-specific parameters.Source: §A.1 · p. 10

A.2 Semantic Graph-to-State Interface

  • Each node’s decision state pairs graph-level context with seven semantically grouped percentiles, serialized in a fixed order for non-generative option-logit readout.The frozen model receives a structured query with anomalous and ordinary options; neither the backbone nor interface uses GAD-specific gradients.Source: §A.2 · pp. 10–11

A.3 Statistical Scoring and Rank Fusion

  • Four numerical scorers operate on the 16-dimensional percentile matrix, and their individual percentile ranks are averaged equally with the decision rank.The final score gives the pretrained decision rank a default weight of 1/5.Source: §A.3 · p. 11

B.1 Dataset Details

Evaluation uses seven target graphs from citation, social, and co-review domains, with anomaly labels reserved for metric computation. The comparisons include trained, generalist, unsupervised, and training-free baselines.

B.1 Dataset Details

  • The seven target benchmarks cover citation, social, and co-review graphs, and their anomaly labels are used only to compute evaluation metrics.The citation and BlogCatalog versions include injected anomalies; Facebook, Weibo, and Amazon use annotated anomalous accounts, behaviors, or suspicious review patterns.Source: §B.1 · p. 11, Table 4 · p. 11

B.2 Baseline Details

  • Comparisons span 15 baselines across supervised, unsupervised, generalist, and training-free GAD, with source-trained methods evaluated on disjoint target graphs.ARC (10-shot) is a label-assisted reference; TFGAD and FreeGAD are reproduced on the evaluation graph versions without target-specific anomaly training.Source: §B.2 · pp. 11–12

B.3 Implementation Details

The appendix reports implementation settings and analyses of decision-state size, anomaly retrieval, and alternative fifth scorers. These analyses compare semantic decision readout with statistical-only or numerical-score alternatives.

B.3 Implementation Details

  • Decision inference uses frozen Qwen/Qwen3-4B option logits without autoregressive decoding, while preprocessing and statistical scoring run independently per graph on CPU.On one GPU, the reported efficiency measurement is about 120 nodes per second with about 9.6 GiB peak allocated memory.Source: §B.3 · p. 12

C.1 Decision-State Evidence Granularity

  • The 7D state has the highest macro-AUPRC across four diagnostic datasets; expanding from 5D to 7D improves all four, while 10D has mixed effects.The 10D state adds gains on Amazon and BlogCatalog but reduces performance on CiteSeer and Cora.Source: §C.1 · p. 12, Table 5 · p. 12

C.2 Cumulative Anomaly Retrieval

  • 122 true anomalies are retrieved among CiteSeer’s top 150 nodes by Jev-GAD, versus 105 by Statistical-only scoring.This is 17 additional detections in the cumulative retrieval comparison.Source: Figure 7 · p. 12, §C.2 · p. 12

C.3 Alternative Fifth-Scorer Controls

  • 59.72% macro-AUPRC for pretrained semantic decisions exceeds Statistical-only (51.41%) and all three numerical fifth-scorer alternatives.All five-score configurations use equal-weight rank fusion; the comparison covers four diagnostic datasets.Source: Table 6 · p. 13, §C.3 · p. 13

C.4 Conditional Ranking Beyond Statistical Scores

Pretrained semantic decisions are evaluated for ranking nodes within groups of similar Statistical-only scores. The reported comparison is restricted to bins containing both anomalous and normal nodes.

C.4 Conditional Ranking Beyond Statistical Scores

  • Nodes are grouped into ten equal-frequency bins by Statistical-only scores, and pairwise ranking AUC is aggregated over bins containing both anomaly classes.Tied scores stay in the same bin, and aggregation is weighted by the number of valid anomaly–normal pairs.Source: §C.4 · p. 13
  • 0.733 macro within-bin AUC for pretrained semantic decisions exceeds Statistical-only (0.684) and Mean-7D (0.564), Max-7D (0.565), and LOF-7D (0.514).The improvements are most pronounced on CiteSeer and Cora.Source: Table 7 · p. 13, §C.4 · p. 13
  • The within-bin analysis covers only 6.68%–8.99% of all anomaly–normal pairs across the four graphs.Its comparison therefore concerns a limited subset of pairs rather than all anomaly–normal rankings.Source: §C.4 · p. 13

Related papers

Loading 2610.30777v1…