{"id":"058faa09-b658-4bc2-9dcc-cfb94e890080","arxiv_id":"2506.00594","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GEL detects anomalous nodes by scoring evidential uncertainty from feature and topology reconstruction, reporting gains on four of five benchmark datasets.","lead":"A new graph anomaly detection method, GEL, replaces plain reconstruction error with evidential uncertainty estimates from Normal-Inverse-Gamma and Beta models of node features and edges. The authors report improved detection accuracy and robustness on several benchmark graphs, though some headline numbers do not match their own tables.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncertainty terms are never isolated: modality ablations drop reconstruction signal too, and topology 'graph uncertainty' (Eq. 8) reduces to 1−2/S−|2Â−1|, so the claim that uncertainty adds signal beyond reconstruction error is unverified.","rationale":"I agree with the reader's weakest assumption, and the manuscript itself provides no counter-evidence: the only uncertainty-relevant experiment (Section 4.5) compares GEL to GCNAE, and the ablation study removes modalities, not uncertainty terms. My analytical check sharpens the concern: after substituting the definitions in Eqs. 6–8, the topology uncertainty pair is equivalent to (Â, S), so it adds at most one scalar per edge beyond the edge probability already used in the reconstruction-error term. This does not prove redundancy—it identifies the precise quantity whose anomaly signal must be demonstrated. Separately, the claim of a 4.64% average AUC improvement is not reproducible from Table 1 (GEL loses to G3AD on Weibo by ~5.8 AUC, and the mean gain over the best baseline on the remaining four datasets is ~2.2), which is a further reason the empirical headline needs scrutiny. These issues are addressable with the released code and a fixed-backbone ablation, so the correct disposition remains CONDITIONAL. If the proposed ablation shows no benefit, GEL should be reframed as a robust GAE variant rather than a paradigm shift from reconstruction error to uncertainty.","tokens_in":23494,"tokens_out":11464,"duration_ms":112946,"concrete_test":"Using the released code, train GEL once per dataset (Weibo, Reddit) and compute AUC/Recall@K plus the Fig. 6 noise/dropout curves for two scorings of the same trained model: the full §4.4 score, and the same score with λ_g=λ_r=0 (reconstruction error only). Run at least 5 seeds and report per-dataset margins. If the full score is not significantly better than reconstruction-only on the same model, or if the uncertainty terms do not flatten the robustness curves, the central claim fails; if it is better, the reader's conditional concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.3 ablates whole modalities (w/o feature, w/o topology); each variant removes the modality's reconstruction error together with its uncertainty terms, so the experiment cannot show that U_f_graph, U_f_reconst, U_t_graph, or U_t_reconst contribute anything beyond reconstruction error. Section 4.5's 'hypothesis' experiment compares GEL against a different model (GCNAE), conflating model change with uncertainty scoring. The concern is concrete: with S=ε+ε̄, Â=ε/S, b=(ε−1)/S, b̄=(ε̄−1)/S, Eq. 8 gives U_t_graph = 1−2/S−|2Â−1|, and U_t_reconst=1/S, so the two topology uncertainty terms are fully determined by the same Beta parameters that already define the edge reconstruction; the only extra information beyond  is the total evidence S. The feature uncertainties β/(α−1) and β/(ν(α−1)) are likewise extra scalars from the same evidence head that outputs γ = X̂. Whether those extra scalars carry anomaly signal is exactly what is never isolated. Without a fixed-backbone uncertainty on/off ablation, the central conceptual claim (uncertainty, not reconstruction error, drives GEL's robustness) is not established, even if Tables 1/2 rankings are reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Graph Evidential Learning (GEL), an unsupervised graph anomaly detection method. GEL replaces the point-estimate decoders of a graph autoencoder with evidential heads: a Normal-Inverse-Gamma (NIG) head for continuous node features and a Beta head for binary topology. The learned hyperparameters yield two uncertainty measures per modality (graph uncertainty and reconstruction uncertainty), and the anomaly score in Section 4.4 is a weighted sum of the four uncertainty terms plus feature and topology reconstruction errors. The authors report state-of-the-art AUC and Recall@K on four of five benchmark datasets, a robustness study under feature noise and edge dropout, and a large-scale DGraphFin experiment.","tokens_in":23843,"tokens_out":7396,"duration_ms":61185,"significance":"If the uncertainty terms indeed carry anomaly signal beyond reconstruction error, GEL would be a practical contribution: it obtains uncertainty in a single forward pass, adds negligible overhead to a standard GAE, and its code is released. The robustness experiments (Figure 6) and the DGraphFin result are useful evidence of scalability. However, the central conceptual claim, that shifting from reconstruction error to uncertainty drives the gains, is not isolated by any experiment. The uncertainty terms are functions of the same evidence parameters that define the reconstruction, and all ablations remove whole modalities or change the model, so the added value of the uncertainty terms remains unverified. The significance of the paper therefore hinges on a fixed-backbone uncertainty on/off ablation that is currently missing.","major_comments":[{"comment":"The experiments do not isolate the uncertainty terms. The ablation variants 'w/o feature' and 'w/o topology' remove the entire modality, dropping that modality's reconstruction-error term from the anomaly score along with its uncertainty terms; the performance drop could therefore be due to loss of reconstruction signal alone. The 'hypothesis' experiment in Section 4.5 compares GEL with GCNAE, which differs in architecture, loss, and training perturbations, so it cannot attribute the improvement to uncertainty. The central claim that uncertainty, not reconstruction error, drives GEL's performance and robustness requires a fixed-backbone ablation: train the GEL evidential model once, then score anomalies with (a) reconstruction error only (lambda_g=lambda_r=0), (b) uncertainty terms only (dropping the reconstruction-error terms), and (c) the full GEL score, and report AUC/Recall@K and the robustness curves.","section":"Section 5.3 and Section 4.5"},{"comment":"The uncertainty terms are deterministic functions of the same evidence parameters that define the reconstruction. For topology, using S=epsilon+epsilon_bar, A_hat=epsilon/S, b=(epsilon-1)/S, b_bar=(epsilon_bar-1)/S, Eq. (8) reduces to U_t_graph = 1 - 2/S - |2*A_hat - 1|, and U_t_reconst = 1/S; both are functions of A_hat (the edge reconstruction probability) and total evidence S. For features, U_f_graph = beta/(alpha-1) and U_f_reconst = beta/(nu*(alpha-1)) are extra scalars produced by the same head that outputs gamma = X_hat. Thus the only information the uncertainty terms carry beyond the reconstruction predictions is a small set of evidence-strength scalars. Because the anomaly score in Section 4.4 always sums these terms with reconstruction error, the paper does not demonstrate that these scalars contain anomaly signal. Please add a statistical comparison (e.g., correlation with anomaly labels) or an ablation that replaces the uncertainty terms with constants or shuffled values while keeping reconstruction error fixed, to show they contribute beyond reconstruction error.","section":"Section 4.1, Eq. (8), and the anomaly score in Section 4.4"},{"comment":"The reported 'average AUC improvement of 4.64%' over the best baseline cannot be reproduced from Table 1. In Table 1, GEL is below the best baseline on Weibo (G3AD 95.1 vs. GEL 89.3); on Reddit, Disney, Books, and Enron the margins over the best baseline are approximately 0.7, 1.5, 5.1, and 1.5 AUC points, respectively, which average to about 2.2 points (or about 1.7 if Weibo is counted as a negative margin). Please state the exact baseline and computation underlying the 4.64% figure, or correct the claim.","section":"Section 5.2, Table 1"}],"minor_comments":[{"comment":"The anomaly score equation is not numbered, although the discussion in Section 4.5 refers to the score in a way that suggests an equation number; please number it consistently and update any cross-references.","section":"Section 4.4"},{"comment":"The caption of Figure 5 appears garbled ('Impact of removing Different Modality.'); please rewrite it to clearly describe the ablation variants and the plotted metric.","section":"Figure 5"},{"comment":"The appendix title contains a typo: 'Netowrk Implemetation' should be 'Network Implementation'.","section":"Appendix D"},{"comment":"There is a typo in 'the total loss defined as::' (double colon); please correct it.","section":"Section 4.3.4"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a data mining venue and is generally clearly written. My main concern for the editor is that the headline 'uncertainty, not reconstruction error' claim is currently supported only by comparisons that conflate model changes with uncertainty scoring; if the requested isolation ablation fails, the contribution reduces to a reparameterized GAE with extra hyperparameters. The 4.64% average improvement figure also needs correction or a precise derivation before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: GEL is a clean adaptation of evidential learning to unsupervised graph anomaly detection, with a solid experimental package, but the paper's headline claim—that uncertainty, not reconstruction error, drives the gains—is not actually tested. The ablations remove whole modalities, so they can't separate uncertainty from reconstruction loss. That's the main thing you should know.\n\nWhat's new: jointly modeling feature reconstruction with a Normal-Inverse-Gamma and topology with a Beta, defining four uncertainty terms, and combining them with reconstruction error in the anomaly score. The integration is neat and the math checks out. The authors also deserve credit for benchmarking on five datasets plus DGraphFin at 3.7M nodes, reporting training/inference time and memory, and releasing code. The robustness study under feature noise and edge dropout is a nice addition.\n\nSoft spots: the key conceptual contribution is unverified. In the score (Eq 17), uncertainty terms always ride alongside reconstruction error. The ablation removes the feature evidence network or the topology evidence network, which removes both the uncertainty terms and the corresponding reconstruction error for that modality. So you learn that both modalities matter, not that uncertainty matters. Section 4.5 compares GEL to GCNAE, a different architecture, so that doesn't isolate uncertainty either. The stress-test math is right: U_t_graph and U_t_reconst are fully determined by the Beta parameters that already give the edge probability Â; the only extra information is total evidence S. Feature uncertainties are extra scalars from the same NIG head that outputs the reconstruction γ. Those scalars could carry signal, but the paper never isolates them. Also, the claimed average 4.64% AUC improvement doesn't follow from Table 1 (GEL loses on Weibo by ~6 points), and the per-dataset tuning of λf/λt/λg/λr isn't disclosed—if labels were used for selection, that's leakage.\n\nThat said, the paper is not misleading in a cynical way. The method is coherent, the limitations section is honest, and the central problem is an experimental design gap, not a mathematical error. The fix is straightforward: run a fixed-backbone ablation that toggles the uncertainty terms on/off in the score, and report tuning protocol.\n\nWho should read it: anyone working on reconstruction-based anomaly detection or evidential learning in graphs. It deserves a serious referee; with the right ablations it could be a solid paper. I'd send it to review, but require the isolation experiment.","headline":"GEL is a clean evidential-learning adaptation for graph anomaly detection with a solid experimental package, but the paper's central claim—that uncertainty, not reconstruction error, drives the gains—is not actually tested by its ablations.","tokens_in":24350,"tokens_out":2819,"would_cite":true,"duration_ms":26321,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper argues that scoring graph nodes by reconstruction uncertainty, derived from evidential distributions, beats scoring them by reconstruction error alone.","keywords":["graph anomaly detection","evidential learning","uncertainty quantification","graph autoencoder","graph neural networks","Normal Inverse-Gamma distribution","Beta distribution","reconstruction error"],"falsifier":"Keep the GEL reconstruction backbone fixed and set the uncertainty weights in the Section 4.4 anomaly score to zero, scoring purely by reconstruction error; if AUC on the five datasets stays essentially unchanged, the claim that uncertainty adds signal is false.","tokens_in":23309,"feed_emoji":"🕸️","tokens_out":6928,"duration_ms":55277,"temperature":0.7,"pith_summary":"This paper claims that unsupervised graph anomaly detection should stop scoring nodes by how badly they reconstruct and instead score them by how uncertain the reconstruction is. It proposes Graph Evidential Learning (GEL), which wraps the feature and topology decoders of a graph autoencoder in higher-order evidential distributions—a Normal Inverse-Gamma for continuous features and a Beta for discrete edges—so that every reconstruction carries two uncertainties: graph uncertainty from conflicting evidence and reconstruction uncertainty from lack of evidence. On five public graphs and the million-node DGraphFin graph, GEL reports state-of-the-art AUC and Recall@K and degrades more slowly than baselines when features are nosified or edges are dropped. The payoff, if true, is that anomaly detectors become less brittle to noise and less prone to overfitting anomalies into normal reconstructions.","feed_headline":"Uncertainty, not reconstruction error, flags graph anomalies","feed_subtitle":"Graph evidential learning lifts anomaly detection on five datasets and stays robust under noise and edge dropout.","key_machinery":"The machinery is the higher-order evidential distribution used as a decoder. For node features, a Normal Inverse-Gamma (NIG) distribution places a Gaussian prior on the mean and an Inverse-Gamma prior on the variance, yielding reconstruction uncertainty $U_{reconst}^f = \\frac{\\beta}{\\nu(\\alpha-1)}$ and graph uncertainty $U_{graph}^f = \\frac{\\beta}{\\alpha-1}$. For topology, a Beta distribution over edge probability with strength $S_{ij} = \\varepsilon_{ij} + \\bar{\\varepsilon}_{ij}$ yields reconstruction uncertainty $U_{reconst}^t = 1/S_{ij}$ and a graph uncertainty based on the balance of supporting and opposing belief masses. The anomaly score defined in Section 4.4 combines these four uncertainty terms with feature and topology reconstruction error, so the same learned evidence drives both the reconstruction and the uncertainty attached to it.","core_discovery":"The central claim is that shifting graph anomaly detection from reconstruction error to uncertainty modeling improves both accuracy and robustness. GEL models the reconstruction of node features and topology as the expected values of two higher-order evidential distributions, a Normal Inverse-Gamma for features and a Beta for edges, trained by a multi-task loss that maximizes reconstruction fit while suppressing evidence where the model errs. Anomaly scores combine reconstruction error with four evidential terms: feature and topology reconstruction uncertainty plus feature and topology graph uncertainty. Experiments report state-of-the-art results on four of the five datasets and on DGraphFin, and the Section 4.5 Cora analysis shows that anomalous classes receive higher normalized uncertainty than normal classes, in contrast to the GCNAE baseline.","pith_inferences":["A sharper ablation than the paper's modality removal would keep the same reconstruction backbone and zero out only the four uncertainty terms in the anomaly score; if accuracy holds, the conceptual claim would be weakened to 'a better graph autoencoder,' and that test is not reported in the paper.","Since the optimal balance between graph and reconstruction uncertainty varies by dataset, an adaptive or learned weighting of the four terms could extend GEL without new labeled data.","The topology graph-uncertainty term measures conflict between supporting and opposing edge evidence, so it might stand alone as a structural anomaly score and connect to local evidence-conflict measures in other graph tasks."],"forward_implications":["Uncertainty-based scoring beats reconstruction-error scoring on attributed graphs with low anomaly rates, where GEL reports its largest gains (Reddit and Books).","GEL's performance degrades more gently than GAE baselines as feature noise and edge dropout increase, so evidential uncertainty is claimed to be the source of that robustness.","The framework scales to a 3.7-million-node graph with about 17% more training time than GAD-NR, making the uncertainty machinery cheap enough for large-scale deployment.","Because the Cora experiment shows anomalous classes getting higher normalized uncertainty than normal classes, the uncertainty terms themselves carry anomaly signal, not just the reconstruction error."],"supporting_citations":[{"why":"Supplies the Normal Inverse-Gamma evidential regression formulation that GEL adapts for feature reconstruction.","marker":"[3]"},{"why":"Establishes evidential learning with higher-order distributions, the conceptual basis for deriving uncertainty from evidence in a single forward pass.","marker":"[59]"},{"why":"Provides Subjective Logic as the formalism for quantifying belief and uncertainty over discrete edge outcomes.","marker":"[30]"},{"why":"Shows the Beta-distribution parameterization of evidence (CEDL) that GEL uses for topology reconstruction uncertainty.","marker":"[2]"},{"why":"Provides the graph autoencoder reconstruction backbone (VGAE) that GEL extends with evidential decoders.","marker":"[33]"},{"why":"The strongest GAE baseline in the comparison, whose neighborhood-reconstruction method GEL is measured against and outperforms.","marker":"[55]"},{"why":"Supplies the nineteen-baseline benchmark setup (BOND) used for the experimental comparisons.","marker":"[42]"}],"fun_headline_variants":["Uncertainty beats reconstruction for graph anomaly detection","How to spot graph anomalies: trust uncertainty","Evidential uncertainty: better graph anomaly detection","Graph anomalies: uncertainty is the new signal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the evidential uncertainty terms reveal anomalies that reconstruction error misses, rather than simply re-encoding the same information that the reconstruction error already contains.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty beats reconstruction for graph anomaly detection","How to spot graph anomalies: trust uncertainty","Evidential uncertainty: better graph anomaly detection","Graph anomalies: uncertainty is the new signal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000669,"raw_usage":{"total_tokens":2985,"prompt_tokens":817,"completion_tokens":2168,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":433,"completion_tokens_details":{"reasoning_tokens":2112}},"tokens_in":433,"tokens_out":2168,"duration_ms":16581,"temperature":1.0,"reasoning_tokens":2112,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:02:35.646494+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Keep the GEL reconstruction backbone fixed and set the uncertainty weights in the Section 4.4 anomaly score to zero, scoring purely by reconstruction error; if AUC on the five datasets stays essentially unchanged, the claim that uncertainty adds signal is false.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Normal Inverse-Gamma evidential regression formulation that GEL adapts for feature reconstruction."},{"cited_title":"2016.Subjective Logic - A Formalism for Reasoning Under Uncer- tainty","cited_arxiv_id":null,"evidence_quote":"Provides Subjective Logic as the formalism for quantifying belief and uncertainty over discrete edge outcomes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows the Beta-distribution parameterization of evidence (CEDL) that GEL uses for topology reconstruction uncertainty."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The strongest GAE baseline in the comparison, whose neighborhood-reconstruction method GEL is measured against and outperforms."},{"cited_title":"Chen, Zhihao Jia, and Philip S","cited_arxiv_id":null,"evidence_quote":"Supplies the nineteen-baseline benchmark setup (BOND) used for the experimental comparisons."}],"review_version":1}