acceptodds
Under review as a conference paper at ICLR 2027

MSS: Multi-Semantic Spaces as a Declared Evidence Layer for Graph Anomaly Detection

Abstract

Benchmarks for graph anomaly detection (GAD) usually change the evidence construction and the consuming model together, so a reported margin does not show which of the two produced it, and an aggregate score hides the view a prediction rests on. Holding the read-out fixed locates the gain in the construction rather than in the model that consumes it: on the same tree head, complete evidence improves the ten-graph mean AUPRC over its baseline feed by 1.65 points under full supervision and 4.17 under the 20+80 budget, and on the largest graph adding regional references to that feed moves AUPRC from 41.12 to 72.64. We propose Multi-Semantic Spaces (MSS), a framework that treats raw and graph-derived feature blocks as materials declared by explicit mechanism hypotheses. A frozen topology backbone, soft regional occupancy, and regional references organize these materials into positions and deviations from a declared comparison unit. MSS-T fits boosted trees to a fixed evidence table, while MSS-N learns private coordinates and per-space experts, updates its references during training, and retains per-space states through stop-gradient fusion. Complete evidence also improves a second library-default head. On ten public benchmarks, using the released GADBench splits and a common ten-seed list, one configuration shared across all ten datasets and never retuned attains the highest mean AUPRC under both budgets. It exceeds the strongest baseline by 2.57 points under full supervision and 4.69 under the 20+80 budget. MSS-N exceeds MSS-T on two fully supervised graphs and exposes coordinate-level diagnostics, so the gains stay attributable and the states inspectable.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph anomaly detection, fraud detection, supervised anomaly detection, semi-supervised anomaly detection, learning on graphs, multi-semantic spaces, declared evidence layer, benchmark evaluation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_pUJ2OpenReview ↗
Rating
3(−1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes MSS, an evidence-construction framework for graph anomaly detection. Instead of changing only the prediction model, it organizes graph-derived features into mechanism-specific spaces and compares nodes against regional references. Two implementations are provided: MSS-T with boosted trees and MSS-N with learned per-space experts. Experiments on ten benchmarks evaluate prediction accuracy, evidence attribution, and scalability.

Critical strengths
  1. The experimental design is a major strength. By holding the prediction head fixed, the authors provide direct evidence that the proposed feature construction contributes to performance rather than attributing all improvements to a new model architecture.

  2. The evaluation is extensive and well controlled. MSS-T achieves the highest mean AUPRC under both full supervision and limited labels. Results on million-node graphs further support the scalability of the approach.

Critical weaknesses
  1. The practical gains are uneven across datasets. Section 5.3 shows that much of the improvement from regional references comes from T-Social, while the gains on several other graphs are small. The paper should make clearer when regional references are beneficial and when simpler graph features are sufficient.

  2. The mechanism-specific interpretation remains stronger than the available evidence. The spaces are named according to anomaly hypotheses, but the reported diagnostic states do not by themselves establish that the corresponding real-world mechanisms have been identified. The paper should distinguish interpretable feature organization from validated explanations.

Questions for authors
  1. What graph characteristics determine whether regional references provide a meaningful advantage over the baseline evidence?

  2. What evidence supports interpreting the learned space-specific outputs as particular anomaly mechanisms?

Reviewer_mFSROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

This paper investigates how to effectively identify distinct types of anomalies. The authors point out that existing approaches often optimize feature construction strategies and model architectures simultaneously, making it difficult to discern the true source of performance gains. Furthermore, these methods typically yield only a final anomaly score, offering little insight into the specific contribution of different feature perspectives to the final prediction. To address these challenges, the authors propose Multi-Semantic Spaces (MSS). This framework constructs multiple semantic spaces—each corresponding to a specific anomaly mechanism—and utilizes a soft-assignment mechanism to derive "regional reference representations." Anomalies are then detected by measuring the deviation of individual nodes from these reference representations. Building on this, the authors develop two variants: MSS-T (based on tree models) and MSS-N (based on neural networks), designed to validate the efficacy of the feature construction strategy and provide predictive insights across various semantic spaces, respectively. The effectiveness of the proposed approach is empirically validated on ten public datasets.

Critical strengths
  1. This paper focuses on how to effectively identify differentiated abnormal patterns, and its research topic has good practical significance.
  2. The proposed Multi-Semantic Spaces explicitly organizes different feature perspectives, and identify anomalous node behaviors through regional reference representations. The overall framework design is intuitive.
  3. The authors conducted experiments under both fully supervised and limited-label settings across ten datasets. The results demonstrate that the proposed method achieves competitive performance in terms of average AUPRC.
Critical weaknesses
  1. The article is poorly readable, containing many custom terms without sufficient explanation, such as evidence construction and the consuming model.
  2. While the paper explicitly models multiple feature perspectives, it lacks a thorough discussion on the rationale behind the selection of these perspectives. It remains unclear whether introducing more semantic space can further improve detection performance, and how to determine whether the existing semantic space is sufficient to cover different anomaly patterns.
  3. The proposed framework constructs richer anomaly detection evidence by modeling multiple semantic spaces, but this may also bring additional computational overhead. Although the authors reported the runtime and memory consumption, there is a lack of systematic computational efficiency comparisons with existing methods in the same experimental settings, so it is unclear whether the performance gains are sufficient to offset the additional computational overhead.
Questions for authors

Please see the Weaknesses.

Reviewer_wW6uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary
  1. The paper proposes MSS, which organizes raw and graph-derived features into mechanism-specific semantic spaces and compares nodes against regional references.
  2. MSS provides two read-outs: MSS-T uses boosted trees on a fixed evidence table, while MSS-N learns space-specific coordinates and experts with gated fusion.
  3. The method is evaluated on ten GADBench datasets under full-supervision and limited-label settings, using common splits and ten random seeds.
  4. MSS-T achieves mean AUPRC scores of 63.77 and 47.86 under the two settings, outperforming the strongest reported baselines by 2.57 and 4.69 points, respectively.
  5. The paper includes fixed-head attribution experiments, component ablations, large-graph evaluations, and per-space diagnostic visualizations.
Critical strengths
  1. The fixed-tree-head experiments help separate the contribution of evidence construction from that of the downstream classifier.
  2. The evaluation covers ten graphs with substantially different structures and anomaly distributions.
  3. The use of common splits and ten random seeds provides a reasonably consistent basis for comparison.
  4. The two read-outs test whether the proposed evidence construction is useful for both tree-based and neural detectors.
  5. The large-graph experiments and per-space outputs provide useful evidence about scalability and model inspection.
Critical weaknesses
  1. The performance gains are uneven across datasets, with T-Social, Questions, and Tolokers accounting for a large share of the reported improvements.
  2. MSS-T falls 1.21 AUPRC points below the strongest baseline on Elliptic, indicating that the method is less effective under at least one temporal evaluation setting.
  3. The semantic spaces depend on manually specified anomaly mechanisms and feature materials, and their construction may require additional design for new graph types.
  4. The diagnostic coordinates are presented as interpretable states, but their usefulness for identifying real anomaly mechanisms is not independently validated.
  5. The evaluation follows a transductive protocol, so the results do not establish performance on entirely unseen graphs.
Questions for authors
  1. Can the authors explain why the gains are concentrated on a few datasets and identify the graph characteristics responsible for these differences?
  2. How much manual effort is required to define the semantic spaces and feature materials for a new application domain?
  3. Can the diagnostic states be validated against known anomaly mechanisms or expert-provided annotations?
  4. How does MSS perform when evaluated on unseen graphs rather than new nodes within an observed graph?
  5. Can the authors clarify which anomaly types or graph structures cause the largest performance degradation, particularly on Elliptic?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.