{"id":"426b36ff-5b32-489d-825b-34a9b1a22206","arxiv_id":"2507.14677","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"AD-GCL improves graph anomaly detection on low-degree tail nodes by pruning head-node edges to forge tail-like views and completing tail neighborhoods using the model's own anomaly scores.","lead":"This paper proposes AD-GCL, a graph contrastive learning method for unsupervised anomaly detection that targets low-degree tail nodes, which existing methods miss. It matters for detecting fraudsters and bots in power-law networks, where most nodes have very few connections.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Early self-scores drive the completion module, but the paper never tests whether those scores are accurate enough or whether the feedback loop amplifies errors.","rationale":"The central claim is empirical: AD-GCL improves unsupervised graph anomaly detection under structural imbalance, especially for tail nodes. The decisive component is anomaly-guided neighbor completion, the only part of the method that uses the model's own outputs to select training data. For that component to work as claimed, the discriminator's early anomaly scores must carry enough signal to choose useful auxiliary nodes, and the resulting feedback must not diverge. The paper provides no convergence or error analysis, and the w=0 ablation is indirect: removing the anomaly term removes the whole term, not just early noise. The proposed freezing test directly varies the amount of corrective information available to p_ano, thereby separating 'guidance helps when accurate' from 'guidance is robust when noisy'. The underspecified φ reinforces the need for precision in Eq. (2), but I do not treat it as a separate fatal flaw because the released code may resolve it. Table 1 and the ablation study are real empirical support, and I see no reason to reject the paper; the correct disposition remains conditional pending the test.","tokens_in":22579,"tokens_out":7711,"duration_ms":104405,"concrete_test":"On Cora and Citeseer, compare the published w=5 setting with a variant in which S is frozen after the first w epochs, so p_ano never receives post-w scores, while keeping all other hyperparameters fixed. If tail AUC differs by less than 1 point, early self-score noise is not harmful and the feedback-loop concern does not land; if tail AUC drops by more than 1 point, the sliding-window update is doing essential corrective work and the paper should provide a sensitivity analysis or a bound on the early-score error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The mechanism claim in Section 'Anomaly-Guided Neighbor Completion' is that the discriminator's own scores, stored over a sliding window of w epochs, define p_ano(u,v) = S_u · S_v^T, which selects auxiliary nodes a whose ego-networks are mixed into tail-node neighborhoods via pmix(·|v) = (1−φ)p(·|v)pnc(·|v) + φ p(·|a)pnc(·|a). The load-bearing premise is that these self-scores are sufficiently accurate during training to identify useful auxiliary nodes. Early in training the discriminator has not converged, so p_ano is approximately noise; because the completed neighborhoods are fed back into the same discriminator through L_intra and L_inter, selection errors can be amplified rather than corrected. The w=0 ablation only shows that removing the anomaly term hurts on Cora; it does not measure early-score quality, and no convergence or error bound is given. Additionally, φ is described only as 'increases with similarity' and 'at most 0.5', with no formula, so Eq. (2) is underspecified; if φ is large, the anchor tail node's own neighborhood is destroyed, and the claimed 'enlargement' would not occur as described. Without a robustness check on p_ano and φ, the central claim that tail-node gains come from anomaly-guided completion rather than from extra augmentation is not fully established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AD-GCL, an unsupervised graph contrastive learning method for anomaly detection that targets structural imbalance. The method has two main components: neighbor pruning, which samples K salient neighbors for head nodes to create 'forged tail nodes' and aligns them with the original head nodes via inter-view contrast, and anomaly-guided neighbor completion, which enlarges the receptive field of tail nodes by mixing their ego networks with those of auxiliary nodes selected using both feature similarity and the model's own anomaly scores. The training objective combines intra-view BCE losses and inter-view InfoNCE losses. Experiments on six citation/bitcoin datasets (plus two in the appendix) report AUC for all, tail, and head nodes, and show that AD-GCL often outperforms ten baselines. The paper also provides ablations, parameter studies, complexity analysis, and a limitations appendix.","tokens_in":22848,"tokens_out":5839,"duration_ms":67898,"significance":"If the reported results are reliable, the paper makes a useful contribution to unsupervised graph anomaly detection by identifying a real weakness of existing GCL methods on low-degree nodes and by demonstrating a practical remedy with consistent tail-node AUC gains. The empirical study is fairly extensive: ten baselines, six main datasets, head/tail breakdowns, ablations, parameter sensitivity, and an appendix with further metrics. The authors also state that source code and datasets are released, which strengthens reproducibility. The main value is the demonstration that structural imbalance can be explicitly addressed within the GCL framework without labels. The core empirical claim is largely supported, though some overstatements and underspecified components need to be corrected.","major_comments":[{"comment":"The claim that 'AD-GCL achieves the best anomaly detection performance on these six datasets' and 'the best AUC scores for both tail nodes and head nodes on most of the datasets' is overstated. In Table 1, on Pubmed, ANEMONE has a higher head AUC (98.20) than AD-GCL (97.88). In Appendix Table 3, on Tolokers, GAD-NR has higher overall AUPRC/AP (30.20/30.21) and higher head AUPRC/AP (31.62/31.63) than AD-GCL (27.02/27.05 and 27.14/27.19, respectively). These exceptions should be explicitly acknowledged, and the abstract/conclusion wording of 'comprehensive superiority' should be tempered to reflect the actual win/loss pattern.","section":"Main Results and Analysis"},{"comment":"The mixing ratio φ in Eq. (2) is not defined. The text says only that φ 'increases with the similarity to the tail node v' and is 'at most 0.5', but no formula or algorithmic specification is given. This makes the neighbor completion step irreproducible and leaves open the possibility that for highly similar auxiliary nodes φ is close to 0.5, which would substantially destroy the anchor node's own neighborhood and contradict the stated goal of enlarging the receptive field. Please provide an exact definition of φ and report the values used in the experiments, ideally with a sensitivity analysis.","section":"Anomaly-Guided Neighbor Completion, Eq. (2)"},{"comment":"The training schedule is ambiguous. The text says that during the initial stage view1 is the original graph and view2 is the neighbor-pruned graph, and that 'in the later stage of training, we use the neighbor completion strategy ... generating two augmented graphs (referred to as view1 and view2)'. It is unclear (a) at what epoch or by what criterion the switch occurs, (b) whether view1 in the later stage remains the original graph or is also the result of completion, and (c) which view(s) are used when computing the anomaly scores S_{v} that feed into p_ano. This underspecification affects both reproducibility and the interpretation of the inter-view contrastive loss.","section":"Graph Contrastive Network"},{"comment":"The evidence that anomaly-guided selection, rather than the mixup augmentation itself, drives the tail-node improvement is incomplete. The distribution p_ano is computed from the discriminator's own scores over a sliding window; early in training these scores are noisy, and because the completed neighborhoods are fed back into the same discriminator, selection errors could be amplified. The w=0 ablation in Figure 4d is only shown for Cora and does not isolate feedback effects. Please add control experiments, for example sampling auxiliary nodes by feature similarity alone (without the p_ano term) and reporting the w=0 ablation on all datasets, or providing an analysis of how p_ano evolves during training.","section":"Anomaly-Guided Neighbor Completion and Parameter Study"}],"minor_comments":[{"comment":"The symbol S is used both for the sliding-window anomaly-score matrix in the completion section and for the final anomaly score in Eq. (10); please use distinct notations to avoid confusion.","section":"Anomaly Score Calculation, Eq. (10)"},{"comment":"The product p(·|u) · psim(·|u) is described as sampling from a multinomial distribution; please state explicitly that the product is normalized to form a valid probability distribution, or define Multinomial to accept unnormalized weights.","section":"Neighbor Pruning Strategy, Eq. (1)"},{"comment":"The numerical degradation percentages (2.46%, 1.87%, 30.62%, 1.89%) reported in the text are not tied to a specific table or figure; please include the underlying values in Figure 3 or a supplementary table for reproducibility.","section":"Ablation Study"},{"comment":"The main text consistently refers to 'six datasets', while the appendix reports results on eight datasets; please harmonize the dataset counts and clarify that Table 3 covers a subset of datasets with AUPRC/AP.","section":"Appendix, Table 3"},{"comment":"The limitations section is a welcome addition; consider also mentioning that the anomaly-guided completion depends on the quality of the discriminator's early training scores, which is currently not analyzed.","section":"Limitations"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the venue and addresses a timely problem. The main empirical finding is credible, but the manuscript overstates its 'comprehensive superiority' and leaves two components (φ and the training schedule) underspecified. These issues are fixable with additional experiments and clarifications. I recommend major revision rather than rejection, as the core approach appears sound and the requested changes are achievable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague —\n\nThe short version: this is a solid, worth-reading empirical paper on unsupervised graph anomaly detection under structural imbalance. The new bits — similarity-aware neighbor pruning that forges head nodes into tail-like views, and anomaly-guided neighbor completion that mixes ego-networks of tail nodes with auxiliary nodes chosen by the model's own scores — are genuinely absent from the earlier degree-bias or GAD literature I know. The head/tail evaluation protocol is a real contribution: the field mostly reports average AUC, which hides exactly the tail-node failure this paper targets.\n\nWhat it does well: the central claim is supported. AD-GCL wins overall and tail AUC on all six main datasets and head AUC on five of six. The ablation is clean: each component contributes, and intra-view contrast dominates. They also report AUPRC/AP, ROC curves, robustness checks (readout, multinomial vs kNN sampling), and two extra datasets (Reddit, Tolokers) in the appendix. That is a serious experimental package.\n\nSoft spots, in order. First, the completion module is a self-bootstrap: the discriminator's own sliding-window scores define the anomaly-similarity distribution that selects auxiliary nodes, and the completed neighborhoods feed back into the same discriminator. There's no analysis of whether early scores are accurate enough, no error bound, no convergence argument. The w=0 ablation shows the guidance helps on Cora, but that's indirect. Also, φ in Eq. (2) is underspecified — 'increases with similarity' and 'at most 0.5' is not a formula, and if φ is large the anchor's own neighborhood is destroyed. Second, 'comprehensive superiority' is overstated: ANEMONE beats AD-GCL on Pubmed head AUC, and GAD-NR has higher overall AUPRC/AP on Tolokers. Third, the name AD-GCL collides with Sadeghi et al.'s adversarial GCL (NeurIPS 2021), which isn't cited — fix that. Fourth, degree-bias baselines (Tail-GNN, GRADE, SAILOR, GRACE) are discussed but not compared; I understand they're semi-supervised node classification, but the paper should say why they can't be adapted.\n\nNone of these are load-bearing. The core claim — tail-node detection improves without hurting head nodes — holds up. The limitations section (homophily, fixed K, no theory) is honest.\n\nBottom line: send it to review. It needs a revision on φ and the self-score loop, but it's a legitimate contribution. I'd cite it if I worked on GAD, and it's a good reading-group topic for discussing self-supervised bootstrapping and evaluation protocols.","headline":"Solid empirical GAD paper with a genuine structural-imbalance angle; the core result holds, but the completion module's self-supervision loop and a few reporting overstatements need fixing.","tokens_in":23396,"tokens_out":7415,"would_cite":true,"duration_ms":58816,"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":"Unsupervised graph anomaly detection can be made to catch low-degree tail anomalies, not just high-degree head anomalies, by forging tail nodes from head-node neighborhoods and completing tail neighborhoods from similar nodes.","keywords":["graph anomaly detection","graph contrastive learning","structural imbalance","tail nodes","neighbor pruning","neighbor completion","self-supervised learning","power-law networks"],"falsifier":"Build or select a benchmark where low-degree normal nodes and low-degree anomalous nodes have nearly identical features but different local connectivity, then compare anomaly-guided completion against completion using feature similarity only or random auxiliary selection; if random or feature-only completion matches AD-GCL's tail AUC, the anomaly-score guidance is not carrying the result, and if performance collapses when the first epochs' scores are frozen, the feedback loop is a liability.","tokens_in":22360,"feed_emoji":"🕸️","tokens_out":8790,"duration_ms":92625,"temperature":0.7,"pith_summary":"This paper argues that existing graph contrastive learning (GCL) methods for anomaly detection are structurally biased: they detect abnormal high-degree (head) nodes well but miss abnormal low-degree (tail) nodes, a failure mode that matters for real power-law networks where fraudsters and impersonators are often low-degree. To fix this, the authors propose AD-GCL, which prunes noisy edges from head nodes to create forged tail nodes, then uses knowledge learned from head nodes to supervise those forged tail nodes and guide detection of genuine tail anomalies. For tail nodes themselves, the model enlarges their small neighborhoods by mixing in ego networks of auxiliary nodes selected by both feature similarity and the model's own anomaly-score similarity. On six benchmark datasets, the paper claims AD-GCL achieves the best AUC overall and on tail and head splits, with the head-to-forged-tail alignment carrying the transfer of detection ability to real tail nodes.","feed_headline":"Graph anomaly detection gets a fix for low-degree nodes","feed_subtitle":"AD-GCL prunes noisy edges on high-degree nodes and completes low-degree neighborhoods, beating ten baselines.","key_machinery":"Two paired augmentation modules drive the method. Neighbor pruning samples $K$ neighbors for a head node $u$ from $\\text{Multinomial}(K; p(\\cdot|u)\\cdot p_{\\text{sim}}(\\cdot|u))$, dropping noisy edges to forge a tail-like node. Anomaly-guided neighbor completion forms $p_{\\text{mix}}(\\cdot|v) = (1-\\varphi)p(\\cdot|v)p_{\\text{nc}}(\\cdot|v) + \\varphi\\, p(\\cdot|a)p_{\\text{nc}}(\\cdot|a)$, where $p_{\\text{nc}} = p_{\\text{sim}}\\cdot p_{\\text{ano}}$ and the anomaly-similarity term $p_{\\text{ano}}(u,v) = S_u \\cdot S_v^{\\top}$ uses discriminator scores from a sliding window of the past $w$ epochs; this mixes a tail node's ego network with the ego network of a similar auxiliary node. These two augmentations generate two contrastive views, and the training objective combines intra-view binary cross-entropy contrast on node-neighbor pairs with inter-view InfoNCE contrast on features and anomaly scores.","core_discovery":"On its own terms, the paper claims that the reason GCL anomaly detectors fail on tail nodes is not insufficient model capacity but a shortage of local structure: low-degree nodes cannot generate enough diverse contrast pairs, so the discriminator overfits their sparse neighborhoods and mistakes them for normal. AD-GCL addresses this with two asymmetric augmentations. For head nodes, it samples K edges using joint neighbor-frequency and feature-saliency information so that a head node is forged into a tail-like node; aligning original head nodes with these forged tail nodes in both feature and score space transfers head-node discriminative knowledge to the tail regime. For genuine tail nodes, it samples auxiliary nodes using the product of feature similarity and an anomaly-similarity score built from a sliding window of discriminator outputs, then mixes the ego networks to enlarge the tail receptive field while preserving degree statistics. The paper reports that this scheme outperforms ten baselines on six datasets, with the largest gains on tail AUC, and that removing either augmentation degrades tail performance.","pith_inferences":["Testable extension: replace the sliding-window self-scores with a momentum encoder or exponentially smoothed scores; if early-training score noise is the limiting factor, tail performance should improve beyond the paper's windowed version.","On heterophilic graphs, the discriminator's neighbor-matching assumption is inverted, so the same pruning and completion recipes would likely need reversed similarity signals; the paper leaves this open.","The method implies a general design rule: structural imbalance in graphs can be attacked by deliberately reshaping degree distributions during augmentation rather than by reweighting the final loss.","A self-adaptive degree threshold could replace the fixed Pareto-based $K$ and extend the gains to graphs with very different degree distributions."],"forward_implications":["Existing GCL anomaly detectors' low tail AUC is a structural-imbalance problem, not just a capacity problem, so methods that ignore degree bias will keep missing low-degree anomalies on power-law graphs.","Head nodes can act as label-free teachers for tail detection: aligning original head nodes with forged tail nodes transfers discriminative knowledge without any anomaly labels.","Anomaly-guided neighbor completion enlarges tail receptive fields while preserving degree statistics, and the ablation study says removing it degrades tail AUC by about 1.9 percent.","The two augmentations and the intra/inter-view losses are complementary; removing intra-view contrast causes the largest degradation, about 30.6 percent on tail nodes, indicating it is the dominant learning signal.","The added modules keep the time complexity at $O(|V|d^2 + |E|d + |V|^2d)$, comparable to existing GCL anomaly detectors."],"supporting_citations":[{"why":"Supplies CoLA, the contrastive baseline whose node-neighbor discriminator and statistical anomaly estimator AD-GCL inherits and extends.","marker":"Liu et al. 2021b"},{"why":"Supplies ANEMONE, the multi-scale contrastive baseline and the motivation for view-level consistency.","marker":"Jin et al. 2021"},{"why":"Supplies GRADATE, the multi-view contrastive baseline whose performance drop on tail nodes motivates the paper.","marker":"Duan et al. 2023"},{"why":"Supplies Tail-GNN's head-to-tail knowledge transfer and the degree-threshold definition of tail versus head nodes used in evaluation.","marker":"Liu, Nguyen, and Fang 2021"},{"why":"Supplies GRADE, the structural-fairness baseline whose degree-proportional pruning AD-GCL's similarity-aware pruning is designed to improve on.","marker":"Wang et al. 2022"},{"why":"Supplies PC-GNN, the label-aware neighbor oversampling method that AD-GCL's unsupervised anomaly-guided completion contrasts with.","marker":"Liu et al. 2021a"},{"why":"Supplies InfoNCE, the inter-view contrastive objective used to align features and anomaly scores between views.","marker":"Van den Oord, Li, and Vinyals 2018"},{"why":"Supplies random walk with restart, the procedure that samples neighbor sets for positive and negative contrast pairs.","marker":"Tong, Faloutsos, and Pan 2006"}],"fun_headline_variants":["AD-GCL targets graph anomaly detection blind spots","Graph contrastive learning's tail-node failure fixed","AD-GCL tackles low-degree anomaly detection","Structural imbalance foils GCL anomaly detectors; AD-GCL fixes it","New GCL framework catches anomalies hiding on low-degree nodes"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline stands on the model's own anomaly scores from the most recent few training epochs being accurate enough to pick which neighborhoods to borrow, even though early in training those scores are noisy and the borrowed neighborhoods feed back into the same scorer.","fun_headline_variants_meta":{"raw":{"variants":["AD-GCL targets graph anomaly detection blind spots","Graph contrastive learning's tail-node failure fixed","AD-GCL tackles low-degree anomaly detection","Structural imbalance foils GCL anomaly detectors; AD-GCL fixes it","New GCL framework catches anomalies hiding on low-degree nodes"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000661,"raw_usage":{"total_tokens":3041,"prompt_tokens":981,"completion_tokens":2060,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":1983}},"tokens_in":597,"tokens_out":2060,"duration_ms":17571,"temperature":1.0,"reasoning_tokens":1983,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:51:42.688004+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Build or select a benchmark where low-degree normal nodes and low-degree anomalous nodes have nearly identical features but different local connectivity, then compare anomaly-guided completion against completion using feature similarity only or random auxiliary selection; if random or feature-only completion matches AD-GCL's tail AUC, the anomaly-score guidance is not carrying the result, and if performance collapses when the first epochs' scores are frozen, the feedback loop is a liability.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies ANEMONE, the multi-scale contrastive baseline and the motivation for view-level consistency."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies GRADATE, the multi-view contrastive baseline whose performance drop on tail nodes motivates the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies GRADE, the structural-fairness baseline whose degree-proportional pruning AD-GCL's similarity-aware pruning is designed to improve on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies InfoNCE, the inter-view contrastive objective used to align features and anomaly scores between views."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies random walk with restart, the procedure that samples neighbor sets for positive and negative contrast pairs."}],"review_version":1}