acceptodds
Under review as a conference paper at ICLR 2027

KU-GAD: Rethinking Semi-Supervised Graph Anomaly Detection with Known–Unlabeled Discrimination

Abstract

Semi-supervised graph anomaly detection (GAD) typically has access to only a small set of verified normal nodes, while anomaly labels remain unavailable. Existing methods infer anomaly information indirectly through normality modeling or surrogate signals, leaving unclear whether the observed unlabeled population can provide direct supervision for anomaly ranking. In this work, we revisit this setting from a population perspective and show that, under a fixed representation and representative normal sampling, the optimal known-unlabeled discrimination score preserves the anomaly-to-normal likelihood-ratio ordering. Based on this insight, we propose KU-GAD, a theory-guided framework for estimating anomaly ranking scores from multi-scale graph representations. Since anomalies may arise from different structural contexts, KU-GAD constructs normal-calibrated multi-hop graph responses to capture anomaly evidence across propagation scales. A hop-sequence Transformer is then employed to adaptively aggregate these responses and produce anomaly scores optimized by the known-unlabeled objective. Extensive experiments on seven graph anomaly detection benchmarks demonstrate the effectiveness of KU-GAD, with further analyses validating the importance of graph context, calibration, and normal-reference quality.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph anomaly detection, semi-supervised learning, anomaly detection, graph representation learning, density-ratio estimation, graph neural networks

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_xzfqOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes KU-GAD, a semi-supervised graph anomaly detection framework that learns anomaly rankings directly by discriminating between verified normal nodes and unlabeled nodes. The authors theoretically show that, under representative normal sampling and a fixed representation, the optimal Known–Unlabeled discriminator preserves the target anomaly ranking. Based on this insight, KU-GAD combines normal-calibrated multi-hop graph representations with a Transformer-based encoder and a Known–Unlabeled discrimination objective. Experiments on seven benchmark datasets demonstrate competitive performance, supported by ablation and sensitivity analyses.

Critical strengths

S1:Clear motivation and theoretical interpretation. The paper provides a clear and intuitive perspective on learning anomaly rankings from known normal and unlabeled nodes, supported by theoretical analysis. S2:Comprehensive experimental evaluation. The paper demonstrates strong performance across multiple benchmarks and includes extensive ablation and sensitivity analyses to validate the proposed method.

Critical weaknesses

W1:Insufficient evaluation of the KU objective. Although the paper provides extensive ablation studies, it lacks a direct comparison of different training objectives under the same encoder. Such experiments would help clarify whether the performance improvements mainly come from the proposed Known–Unlabeled discrimination objective or the representation architecture. W2:Fairness of baseline comparisons. Some baseline results are taken directly from previous papers rather than reproduced under the same experimental settings. The authors should further clarify whether the data splits, training protocols, and hyperparameter settings are sufficiently consistent to ensure fair comparisons. W3:Lack of computational efficiency analysis. The proposed method involves multi-hop graph propagation and Transformer-based aggregation, which may introduce additional computational overhead. However, the paper lacks a detailed analysis of training time, inference cost, and memory consumption, making it difficult to assess its scalability and practical efficiency.

Questions for authors

Q1. Effectiveness of the KU Objective. Could the authors provide an objective-level ablation study using the same encoder but different training objectives, such as normal-only or pseudo-anomaly-based objectives? This would help clarify whether the performance improvements primarily come from the proposed KU objective or the representation architecture. Q2. Practical Applicability of the Theoretical Assumption. The theoretical analysis assumes that the known normal nodes are representative of the overall normal population. Although the paper investigates reference bias, could the authors further discuss how this assumption can be satisfied in real-world scenarios and how KU-GAD performs when the assumption is violated? Q3. Fairness of Baseline Comparisons. Some baseline results are taken from previous studies. Could the authors clarify whether these methods follow consistent data splits, normal-reference budgets, evaluation metrics, and model-selection protocols? A more controlled comparison would help establish the reliability of the reported improvements.

4 comments
AuthorsOpenReview ↗

A1. Effectiveness of the KU Objective

We sincerely thank the reviewer for highlighting the importance of isolating the contribution of the KU objective. Following this valuable suggestion, we conducted an objective-level ablation with a shared encoder, comparing KU-GAD against Deep SVDD (Ruff et al., ICML 2018) and OC-NN (Chalapathy et al., 2018), two representative normal-only learning methods.

All methods share the same normal-calibrated multi-hop inputs, Transformer encoder architecture, paired initialization, reference splits, and main-objective training budgets, with objective-specific scoring heads.

We further include KU-hinge, which replaces the original KU-logistic loss (Eq. 5) with a unit-margin hinge objective while preserving K/U supervision:

Lhinge=12EK[max⁡(0,1+s)]+12EU[max⁡(0,1−s)]\mathcal{L}_{\mathrm{hinge}} = \frac{1}{2}\mathbb{E}_{K}[\max(0,1+s)] + \frac{1}{2}\mathbb{E}_{U}[\max(0,1-s)]

Objective-level ablation with a shared encoder (mean AUROC / AUPRC, 10 paired seeds):

ObjectiveAmazonPhotoEllipticT-FinanceTolokersQuestions
Deep SVDD-K.729/.200.673/.161.319/.078.702/.124.573/.289.546/.045
Constrained OCNN-K.412/.086.583/.141.435/.091.635/.134.484/.247.509/.035
KU-hinge.872/.381.435/.087.871/.455.836/.359.650/.318.559/.043
KU-GAD (KU-logistic).909/.813.714/.198.911/.779.833/.480.650/.372.566/.060

All methods are evaluated using common fixed training configurations, independently of the label-free selection used in the original benchmark.

KU-logistic consistently surpasses both one-class objectives in mean AUROC and AUPRC across all six datasets. The same advantage holds in an additional experiment where every method receives identical K-only reconstruction pretraining before objective-specific optimization.

By controlling the encoder architecture and initialization, these results provide direct evidence that the proposed KU discrimination objective contributes to the performance improvements beyond the graph representation architecture itself, validating its role as a central component of KU-GAD.

AuthorsOpenReview ↗

A2. Representative Normal Sampling

We appreciate the reviewer's thoughtful question regarding the practical applicability of the representative-normal assumption. We agree that understanding reference representativeness is important for applying KU-GAD in real-world scenarios, and we clarify both the practical considerations and empirical evidence below.

The representative-normal assumption concerns the coverage of the underlying normal population, rather than merely the correctness of known-normal labels.

Practical applicability. In real-world applications, representative references can be constructed by sampling candidate nodes broadly from the target population and subsequently verifying their normality. Ensuring coverage across diverse behavioral, feature, and structural subgroups can further reduce selection bias. This provides a practical reference-acquisition strategy consistent with our theoretical assumptions.

Theoretical interpretation. When the reference distribution differs from the target normal distribution, the population-optimal score decomposes as:

log⁡pU(z)pK(z)=log⁡pU(z)pN(z)+log⁡pN(z)pK(z)\log\frac{p_{U}(z)}{p_{K}(z)}=\log\frac{p_{U}(z)}{p_{N}(z)}+\log\frac{p_{N}(z)}{p_{K}(z)}

The additional term characterizes the effect of reference-selection bias on anomaly ranking.

Empirical evidence. We have directly investigated violations of representativeness in Section 5.4, Figure 2, and Appendix C.3 (Table 7), using controlled reference-bias experiments across six datasets and five paired seeds.

Reference-bias sensitivity (AUROC / AUPRC):

Reference samplingAmazonPhotoEllipticT-Finance
Unbiased (β=0).917/.874.933/.910.897/.886.870/.802
Mild bias (β=0.5).862/.810.903/.878.923/.904.882/.820
Strong bias (β=4).533/.337.544/.337.849/.742.817/.507

The results demonstrate dataset-dependent robustness to reference shift. Under mild bias, Elliptic and T-Finance retain or slightly improve their performance, while stronger bias produces more substantial degradation, particularly on Amazon and Photo. The full six-dataset results in Table 7 confirm that severe bias reduces both metrics across all evaluated graphs.

Our complementary reference-budget analysis (Appendix C.4) further examines supervision levels from 1% to 15%, distinguishing the amount of normal supervision from its representativeness.

Together, the theoretical analysis and controlled experiments establish a practical guideline for representative normal-reference acquisition and an empirical characterization of KU-GAD's robustness under reference shift.

AuthorsOpenReview ↗

A3. Baseline Fairness

We thank the reviewer for emphasizing the importance of fair and consistent baseline comparisons. In response to this valuable concern, we conducted a new matched evaluation of RHO and GraphNC within our unified experimental framework.

We implemented both methods following their published approaches, adapting their data interfaces and known-normal reference handling to the common graph and supervision protocol. For GraphNC, this additionally involved retraining its GGAD teacher on the matched graph data. We preserve each method's own training and checkpoint-selection procedures while matching the graph data, reference splits, and evaluation protocol.

All three methods were evaluated using identical canonical graphs, ten paired known-normal reference splits, the same evaluation populations, and consistent AUROC/AUPRC calculations.

Matched baseline comparison (AUROC / AUPRC, 10 paired splits; RHO and GraphNC adapted to the common protocol):

MethodAmazonPhotoEllipticT-FinanceTolokersQuestions
RHO.837/.438.711/.237.718/.196.488/.049.520/.288.496/.033
GraphNC.666/.268.591/.133.503/.116.724/.245.479/.240.483/.037
KU-GAD (Ours).904/.738.945/.803.911/.783.892/.714.730/.432.583/.059

KU-GAD achieves the highest mean AUROC and AUPRC across all six datasets, demonstrating its effectiveness under matched graph data, known-normal supervision, and evaluation protocols.

These additional controlled results strengthen the empirical evidence for KU-GAD's improvements over representative graph anomaly detection baselines.

AuthorsOpenReview ↗

A4. Computational Efficiency

We sincerely appreciate the reviewer's suggestion to examine computational efficiency and scalability. To directly address this concern, we provide both a complexity analysis and empirical measurements of training time, inference cost, and memory consumption, including evaluation on the million-node-scale DGraph dataset.

Complexity analysis. KU-GAD avoids quadratic attention over graph nodes. Its multi-hop representations are constructed through sparse graph propagation, which can be precomputed for a fixed configuration. For a graph with NN nodes, MM edges, feature dimension FF, and propagation depth HH, propagation and normal-reference calibration require O(H(M+N)F)O(H(M+N)F) operations.

Importantly, the Transformer performs attention across hop representations within each node, rather than across graph nodes. With L=H+2L=H+2 hop-sequence tokens (including the CLS token) and hidden dimension dd, its attention complexity is O(NL2d)O(NL^2d) per layer, with additional linear-in-NN projection and feed-forward costs.

Consequently, for fixed model dimensions and propagation depth, KU-GAD scales as O(N+M)O(N+M) on a sparse graph, rather than O(N2)O(N^2). This architectural property makes multi-hop propagation and Transformer-based aggregation practical for large-scale graphs.

We further measured the practical costs on Photo, T-Finance, and DGraph.

Computational efficiency (single RTX 2080 Ti, fixed configuration):

DatasetPreparationTrainingFull-U inferencePeak GPU memory
Photo16.51 s23.73 s0.566 s0.547 GiB
T-Finance39.90 s25.43 s0.095 s0.854 GiB
DGraph203.40 s238.65 s5.095 s0.346 GiB

Inference uses prepared graph representations; GPU memory denotes peak allocated training memory. The DGraph training measurement includes diagnostic operations.

Notably, on DGraph with approximately 3.7 million nodes, graph preparation takes 203 seconds, fixed-configuration training takes 239 seconds, and inference over 3.519 million unlabeled nodes takes only 5.10 seconds. Peak training GPU allocation is 0.346 GiB, with a peak sampled CPU RSS of 6.711 GiB.

These results demonstrate that KU-GAD's multi-hop propagation and Transformer-based cross-hop aggregation remain computationally practical on large sparse graphs, combining favorable scaling with efficient training and batched inference.

Reviewer_fTgtOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes KU-GAD, a semi-supervised graph anomaly detection framework that directly exploits the distributional contrast between known-normal and unlabeled nodes. The authors show that, under representative normal sampling and a fixed representation, the population-optimal known–unlabeled discriminator preserves the anomaly-to-normal likelihood-ratio ordering. Based on this insight, KU-GAD combines normal-calibrated multi-hop graph responses with a hop-sequence Transformer to learn anomaly scores without anomaly labels. Experiments on seven benchmarks demonstrate strong performance, supported by architectural ablations and analyses of normal-reference quality.

Critical strengths
  1. The known–unlabeled discrimination formulation provides a principled alternative to normality modeling and pseudo-anomaly generation, with a clear population-level interpretation of anomaly ranking. The normal-calibrated multi-hop representations offer an effective graph-specific realization of this principle.

  2. KU-GAD achieves strong performance across seven benchmarks, with substantial improvements on several datasets, particularly in AUPRC. The graph-permutation and reference-selection bias experiments further support the method's design and theoretical motivation.

Critical weaknesses
  1. The population-ranking result builds on established density-ratio estimation principles, limiting its theoretical novelty. The graph-specific methodological contributions could be more clearly distinguished from the existing statistical foundations.

  2. The label-free selection criterion does not always align with anomaly-ranking performance, as the H=0 variant outperforms the selected configuration on Amazon. This raises concerns about its reliability under finite samples.

Questions for authors

Q1. Could the authors further clarify the graph-specific methodological contributions of KU-GAD beyond the established density-ratio interpretation?

Q2. How reliable is the label-free configuration selection when K/U discrimination is weak, as observed on Amazon?

Q3. Could the authors provide a controlled comparison between KU discrimination and alternative normal-only training objectives to further validate the proposed supervision principle?

Q4. How robust is KU-GAD to the choice of cross-hop aggregation mechanism, particularly when replacing the Transformer reader with a lightweight learnable alternative?

5 comments
AuthorsOpenReview ↗

A1. Graph-specific methodological contributions

We thank the reviewer for the insightful questions. We clarify the graph-specific contributions and present additional evidence on label-free configuration selection, KU supervision, and cross-hop aggregation.

While density-ratio estimation provides the statistical foundation of KU supervision, KU-GAD contributes a graph-specific representation framework that combines normal-calibrated multi-hop responses and adaptive cross-hop aggregation.

Graph propagation changes normal-reference statistics across neighborhood scales. KU-GAD addresses this through hop-wise normal calibration, while retaining and aggregating multiple propagation responses to capture anomaly evidence at different structural ranges.

Importantly, the following ablations hold the KU objective fixed, directly isolating the contribution of the graph representation.

Graph representation ablation on Photo (AUROC / AUPRC):

ConfigurationAUROCAUPRC
Without calibration.722.232
Without propagation.672.189
Permuted graph-node associations.590.139
Terminal hop only.899.666
KU-GAD (Ours).949.806

On T-Finance, hop-wise recalibration further improves performance from .703/.094 with input-only calibration to .890/.710 with both calibration stages.

These results establish that meaningful graph structure, normal-relative multi-hop representations, and adaptive aggregation are essential contributors beyond the underlying statistical supervision principle.

AuthorsOpenReview ↗

A2. Reliability of label-free model selection

KU-GAD uses K/U membership discrimination to select configurations without anomaly validation labels. Under representative normal sampling, for a fixed scoring function:

AUCK/U=12+πA(AUROCN/A−12)\mathrm{AUC}_{K/U}=\frac12+\pi_A\left(\mathrm{AUROC}_{N/A}-\frac12\right)

where πAπ_A is the anomaly proportion among unlabeled nodes. This establishes the population-level connection between membership discrimination and anomaly ranking.

Regarding your observation on Amazon, the originally selected H=9 configuration achieves AUROC of .9046, compared with .9329 at H=0. To assess selection reliability beyond this individual case, we conducted an independent reference-resampling study comparing adaptive selection with a predefined fixed configuration.

Label-free selection versus fixed configuration (AUROC):

DatasetLabel-free selectedFixed H=3, step=32
Amazon.9175.9322
Photo.9059.9307
T-Finance.8604.8221
Questions.5745.5640

These values come from the additional resampling study, separately from the original Amazon example.

Label-free selection improves AUROC on T-Finance and Questions, illustrating the value of adapting configurations to graph-specific characteristics rather than imposing one global depth.

We further examined reference-budget sensitivity. On Amazon, increasing normal-reference coverage from 1% to 5% to 15% improves selected AUROC from .782 to .885 to .918, while reducing depth-selection regret from .081 to .025 to .016.

These findings support the selector as an offline, anomaly-label-free adaptation mechanism. The primary methodological contributions of KU-GAD remain its K/U supervision and normal-calibrated multi-hop representation, which can also be used with predefined configurations.

AuthorsOpenReview ↗

A3. Contribution of KU discrimination

To further validate the proposed K/U supervision principle, we compared KU-GAD against two representative normal-only objectives: Deep SVDD (Ruff et al., ICML 2018) and OC-NN (Chalapathy et al., 2018).

All methods share the same normal-calibrated multi-hop inputs, Transformer encoder architecture, paired initialization, reference splits, and main-objective training budgets, with objective-specific scoring heads.

We also included KU-hinge, a unit-margin counterpart to the original KU-logistic objective (Eq. 5), using the same K/U membership supervision:

Lhinge=12EK[[1+s]+]+12EU[[1−s]+]\mathcal L_{\mathrm{hinge}}=\frac12\mathbb E_K[[1+s]_+]+\frac12\mathbb E_U[[1-s]_+]

Encoder-controlled objective comparison (mean AUROC / AUPRC, 10 paired seeds):

ObjectiveAmazonPhotoEllipticT-FinanceTolokersQuestions
Deep SVDD-K.729/.200.673/.161.319/.078.702/.124.573/.289.546/.045
Constrained OCNN-K.412/.086.583/.141.435/.091.635/.134.484/.247.509/.035
KU-hinge.872/.381.435/.087.871/.455.836/.359.650/.318.559/.043
KU-GAD (KU-logistic).909/.813.714/.198.911/.779.833/.480.650/.372.566/.060

Results use common fixed training configurations rather than the original benchmark's label-free-selected configurations.

KU-logistic achieves higher mean AUROC and AUPRC than both normal-only alternatives on every dataset. This advantage is also consistently observed when all methods share a reconstruction-pretrained encoder initialization.

The results provide direct empirical support for learning from both known-normal and unlabeled populations, demonstrating that the effectiveness of KU discrimination extends beyond the shared graph representation architecture.

Reviewer_6M5hOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

KU-GAD addresses semi-supervised graph anomaly detection using verified normal nodes and an unlabeled target population. Its theoretical basis is that, when representations are fixed and the normal sample is representative, the population-optimal discriminator separating known normal samples from unlabeled samples ranks nodes in the same order as the anomaly-to-normal likelihood ratio. Neither anomaly labels nor an estimate of anomaly prevalence is needed. The model combines feature calibration against a normal reference set, multi-hop graph propagation, and a Transformer that aggregates representations across hops, trained with a balanced logistic objective.

Experiments on seven graph benchmarks measure performance using AUROC and AUPRC, with KU-GAD reporting the strongest results on six datasets. Further analyses examine the roles of calibration, graph context, and aggregation across propagation scales, along with the effects of reference set size and sampling bias. Model configurations are selected using held-out membership in the known normal and unlabeled sets, without consulting anomaly labels.

Critical strengths

1. A clear formulation of supervision for graph anomaly detection with limited verified normals

The paper explains how distinguishing verified normal nodes from an unlabeled population can recover a useful anomaly ranking without individual anomaly labels or an estimate of anomaly prevalence. The density ratio principle has established theoretical roots, but its formulation in this setting offers a useful alternative to generating pseudo-anomalies. The paper also separates ranking equivalence at the population level from guarantees for a model trained on finite samples, clearly defining the scope of its theoretical claims.

2. Strong results supported by informative ablations and failure analyses

KU-GAD reports the highest AUROC and AUPRC on six of seven benchmarks, including substantial AUPRC gains on several datasets. Experiments across five random seeds examine more than component removal. The permuted graph control tests whether propagation gains rely on meaningful graph context. Analyses of calibration, reference set size, and sampling bias probe the method’s central assumptions. Together, these experiments help explain where the approach succeeds and where its performance deteriorates.

Critical weaknesses

1. The novelty lies mainly in the graph implementation rather than the statistical principle

The population ranking result follows from established links between logistic discrimination and density ratio estimation. The encoder combines standardization against normal references, propagation across multiple hops, and Transformer aggregation. Although the paper acknowledges these foundations, it does not compare alternative anomaly learning objectives using the same encoder. The contribution of the known versus unlabeled objective is therefore not clearly separated from that of the representation and calibration pipeline.

2. Performance relies heavily on the number and representativeness of verified normal samples

The reference bias experiments show severe degradation when normal samples are drawn from a narrow subset of the population. The budget study also reveals a substantial gap between the 1% and 15% candidate budget settings. Under the default budget, verified normal nodes account for approximately 11.7% to 14.6% of all nodes across the six datasets evaluated with a common protocol. Verifying this proportion may require considerable effort. These dependencies do not contradict the population result, but they restrict practical use when normal references are scarce or systematically selected.

3. The main baseline comparisons are not fully controlled

Baseline results are taken from RHO and GraphNC rather than reproduced under matching conditions. Appendix B.3 states that the DGraph comparison follows the respective reported protocols, which differ in their treatment of graph direction, background nodes, and the population being scored. These differences are disclosed, but variation in preprocessing, data splits, and configuration selection leaves the exact size of the advantage uncertain. Rerunning the strongest baselines under a common protocol would make the empirical case stronger.

4. The stability of configuration selection without anomaly labels remains unclear

Appendix D reports small differences in membership AUC on the validation set. It also identifies an Amazon checkpoint that improves anomaly detection metrics despite having a lower membership AUC. On Questions, the selection signal is weak, and the chosen propagation depth lies at the boundary of the search range. These findings suggest that the selection metric may not reliably track anomaly ranking quality in finite samples. The paper should quantify selection variability and compare the procedure with a simple fixed configuration.

Questions for authors

1. Contribution of the training objective Can you compare the KU objective with a representative one-class or positive-and-unlabeled learning objective while keeping the calibrated encoder and supervision budget unchanged? This would help determine how much of the gain comes from directly distinguishing the two populations rather than from the representation and calibration pipeline.

2. Comparability of the baselines Can you rerun RHO and GraphNC on representative datasets using identical graph data, reference sets, and evaluation populations? For DGraph, which differences from the published baseline protocols could affect the comparison? A matched evaluation would be useful if feasible.

3. Reference set dependence and model selection How stable are the selected propagation depth and checkpoint across independent development splits, especially when normal reference sets are small or biased? Within each membership validation fold, are calibration statistics computed using only the normal samples in the training portion? Does a fixed configuration achieve comparable anomaly ranking performance? These details would help establish how reliably the proposed selection procedure works without anomaly labels.

4 comments
AuthorsOpenReview ↗

Contribution of KU supervision

To directly assess the contribution of KU supervision, we conducted an encoder-controlled comparison against two established one-class objectives: Deep SVDD (Ruff et al., ICML 2018) and OC-NN (Chalapathy et al., 2018).

All methods share the same normal-calibrated multi-hop inputs, Transformer encoder architecture, paired initialization, reference splits, and main-objective training budgets, with objective-specific scoring heads.

We additionally evaluate KU-hinge, a unit-margin alternative to our original KU-logistic objective (Eq. 5):

Lhinge=12EK[max⁡(0,1+s)]+12EU[max⁡(0,1−s)]\mathcal{L}_{\mathrm{hinge}} = \frac{1}{2}\mathbb{E}_{K}[\max(0,1+s)] + \frac{1}{2}\mathbb{E}_{U}[\max(0,1-s)]

Encoder-controlled objective comparison (mean AUROC / AUPRC, 10 paired seeds):

ObjectiveAmazonPhotoEllipticT-FinanceTolokersQuestions
Deep SVDD-K.729/.200.673/.161.319/.078.702/.124.573/.289.546/.045
Constrained OCNN-K.412/.086.583/.141.435/.091.635/.134.484/.247.509/.035
KU-hinge.872/.381.435/.087.871/.455.836/.359.650/.318.559/.043
KU-GAD (KU-logistic).909/.813.714/.198.911/.779.833/.480.650/.372.566/.060

Results use shared encoder initialization and fixed training configurations, separately from the original label-free-selected benchmark.

KU-logistic consistently outperforms both normal-only objectives in mean AUROC and AUPRC across all six datasets. We further repeated the comparison using a shared K-only reconstruction-pretrained initialization, obtaining the same consistent advantage over both one-class alternatives.

These results directly demonstrate that distinguishing known-normal and unlabeled populations contributes substantially to anomaly detection performance, beyond the shared normal-calibrated graph representation.

AuthorsOpenReview ↗

Baseline comparability

To strengthen baseline comparability, we implemented RHO and GraphNC following their published methods within our unified evaluation framework. We adapted their data interfaces and known-normal reference handling to the common graph and supervision protocol; for GraphNC, this additionally involved retraining its GGAD teacher on the matched graph data.

All three methods were evaluated on identical canonical graphs, ten paired known-normal reference splits, and the same evaluation populations, using consistent AUROC/AUPRC calculations.

Matched baseline comparison (AUROC / AUPRC, 10 paired splits; RHO and GraphNC adapted to the common protocol):

MethodAmazonPhotoEllipticT-FinanceTolokersQuestions
RHO.837/.438.711/.237.718/.196.488/.049.520/.288.496/.033
GraphNC.666/.268.591/.133.503/.116.724/.245.479/.240.483/.037
KU-GAD (Ours).904/.738.945/.803.911/.783.892/.714.730/.432.583/.059

KU-GAD achieves the highest mean AUROC and AUPRC on all six datasets under the matched evaluation, providing additional controlled evidence beyond the literature-reported results in Table 1.

For DGraph, we adapt the original target/background-node protocol to our known-normal/unlabeled setting. Verified normal target nodes (label 0) provide K, whereas anomalous target nodes (label 1) remain unlabeled. Background nodes (labels 2/3) are retained in both graph propagation and U-side training, but excluded from anomaly evaluation. We use a 15% candidate-reference sampling protocol rather than the official train/validation/test masks, and evaluate exclusively on unlabeled target nodes. These protocol differences are detailed in Appendix B.3; accordingly, DGraph is reported separately from the matched six-dataset comparison.

AuthorsOpenReview ↗

Reference dependence and model selection

Under representative normal sampling, the population-level relationship

AUCK/U=12+πA(AUROCN/A−12)\mathrm{AUC}_{K/U}=\frac{1}{2}+\pi_{A}(\mathrm{AUROC}_{N/A}-\frac{1}{2})

provides the theoretical basis for anomaly-label-free configuration selection.

We evaluated the stability of label-free configuration selection across independently sampled normal-reference sets and training seeds, examining both the selected configurations and their resulting anomaly-ranking performance.

Selection stability and performance:

DatasetModal H (frequency)Modal checkpoint (frequency)H-selection regretSelected AUROCFixed H=3, step=32
Amazon0 (28%)32 (42%).0160.9175.9322
Photo3 (22%)32 (78%).0319.9059.9307
T-Finance7 (14%)256 (25%).0130.8604.8221
Questions0 (33%)32 (17%).0272.5745.5640

Configuration stability and ranking robustness are distinct. The exact selected depths vary across reference samples, with modal frequencies of 14–33%; checkpoint selection is more concentrated on Photo (78%) but varies more on the other datasets. Nevertheless, the retrospective H-selection regret remains between .013 and .032 AUROC, indicating a relatively limited ranking penalty from depth selection despite variability in the selected configurations.

The fixed H=3, step=32 configuration provides a competitive alternative, outperforming adaptive selection on Amazon and Photo, while adaptive selection performs better on T-Finance and Questions. Thus, label-free selection provides dataset-dependent adaptation benefits, rather than a uniform advantage over fixed configurations.

Reference coverage affects selection reliability. On Amazon, increasing the candidate normal-reference budget from 1% to 5% to 15% improves selected AUROC from .782 to .885 to .918, while reducing H-selection regret from .081 to .025 to .016. This demonstrates that broader normal-reference coverage improves both anomaly ranking and the reliability of depth selection.

Regarding calibration integrity: within each membership-validation fold, calibration statistics are computed exclusively from training-side known-normal nodes. Held-out validation normals are excluded from calibration fitting and training loss.

Together, these findings characterize the variability of label-free configuration selection, its associated ranking cost, and the role of reference coverage. They also establish a practical fixed-configuration alternative when adaptive selection is unnecessary.

Thank you for your detailed rebuttal and additional experiments! I greatly appreciate that the authors responded to several of my concerns and strengthened the paper’s persuasiveness with supplementary results. My overall assessment remains a weak accept, and I will therefore maintain my original score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.