MV-FGAD: Towards Efficient and Effective Federated Graph Anomaly Detection via Multi-view Learning

Summary

MV-FGAD proposes a federated graph anomaly detection (GAD) framework built from three components: a federated knowledge-learning module that combines FedAvg with a per-client probability-propagation matrix mixing the local adjacency with a knowledge-guided term; a three-view learner (an attribute-only MLP, a structure-aware GCN over the propagation matrix, and a "weak-anomaly" dynamic learner that iterates a relational-similarity signal through an MLP); and a Mahalanobis-distance (MHD) scoring head that averages per-view scores. The conceptual move is to stop treating federated GAD as a neighborhood-aggregation problem — which the authors argue is broken by client-side subgraph partitioning — and to recast it as multi-view mining of anomalies of different strengths, with an explicit view aimed at the "weak" regime where attribute or structural irregularity is small enough to be blurred by aggregation. Relative to the LG-FGAD / FGAD / FedCLGN line, the practical selling points are a lighter shared-parameter footprint and, per the reported measurements, order-of-magnitude runtime savings.

The framing is coherent and the seven-dataset evaluation is broader than most federated GAD papers. However, an independent execution of the method (necessarily reimplemented, since the advertised repository ships no code) does not support the paper's central claims. MV-FGAD's own AUROC reproduces within 2–3 points on four of five main datasets, but the headline Amazon-all AUROC (90.92) collapses to ~70, the reported best average rank (1.0/1.2) becomes mid-pack (~5.0/5.2), and the runtime and OOM advantages invert. Some of these inversions are entangled with baseline-faithfulness limitations (the reimplemented baselines are stronger and lighter than the originals), so they are not all proven paper errors — but the effect is that neither the effectiveness nor the efficiency thesis could be confirmed, and the method needed undocumented numerical fixes just to run. On top of this, several prose claims are contradicted by the paper's own tables, and the method is under-specified in ways that matter for reproduction. The core idea is worth pursuing, but the numerical and prose claims must be brought back into agreement with the evidence and the artifacts released.

Strengths

• Problem framing. Identifying neighborhood aggregation as a poor fit for FL client subgraphs is a clean, well-motivated observation, and Fig. 3 / Fig. 5 give an intuitive handle on why "weak" anomalies matter under partitioning.
• Simple, defensible architecture. Stacking three narrow views (attribute-only MLP, structure-smoothed GCN, relational-similarity iterate) is straightforward and avoids the exotic student–teacher / contrastive machinery of prior work.
• MHD scoring choice. Using Mahalanobis distance rather than a binary classification head is a sensible way to handle severe class imbalance and yields per-view scores that compose naturally.
• Breadth of evaluation. Seven datasets spanning social media, work collaboration, co-review, question answering, and Bitcoin transactions is broad, and the two large-scale sets (Questions, Elliptic) genuinely stress-test scalability.
• Method-internal reproducibility. The independent run does confirm the less headline-grabbing facts: MV-FGAD's absolute AUROC on four of five main datasets is close (within ~2–3 points), its per-round parameter size matches exactly (0.022 MB), and its run-to-run variance is low.

Weaknesses

• Headline effectiveness does not reproduce. The standout Amazon-all AUROC of 90.92 could not be approached by an independent reimplementation (~70), and the reported best average rank (1.0/1.2) came out mid-pack (~5.0/5.2). The large Amazon-all margins over classification baselines (+41.2%/+177.8%) and over LG-FGAD (+18.3%/+96.7%) reversed sign under reproduction. The rank collapse is partly a replication artifact (reimplemented baselines are stronger than the paper's), but the failure to reproduce MV-FGAD's own headline number is not.
• Efficiency thesis does not reproduce. The "0.33% of FedCLGN" runtime claim inverted — MV-FGAD was the slowest of four methods in an independent run — and the FedCLGN-OOM-on-Elliptic observation inverted, with MV-FGAD the most memory-hungry method. Part of this depends on the original heavy baselines being out of scope, but the quadratic $O(\bar n^2 d')$ relation matrix (acknowledged in Appendix A.2) is intrinsic and cuts against the efficiency identity of the paper.
• Overclaims contradicted by the paper's own tables. Three generalisations fail a cross-check: "combining any two views consistently improves performance" (Amazon AAL 75.62 vs. AAL+WADL 68.77 AUROC, Table 2); "Across all datasets, MHD achieves a better balance…" (Table 6: Cosine/Euclidean beat MHD on Questions/Elliptic, Euclidean beats MHD on Elliptic AUROC by ~12 pts); and "multi-view … more efficient and effective across all datasets" (single-WADL wins on Elliptic in Table 5).
• Central WADL argument is intuition-only and unstable. The claim that iteratively accumulating the relational-similarity signal makes "weak signals … no longer be weak" is the pivotal justification but is supported only by density plots on one client, has no analysis of when the recursion amplifies vs. smooths, and gives no explicit update equation for $\tilde{\mathbf{S}}^{(l)}$. Reimplementing it diverged to NaN without undocumented stabilization.
• Notation and specification gaps. The propagation matrix flips between $\tilde{\mathbf{P}}$ and $\mathbf{P}_i$ without a stated relationship (Eq. 7); $l$ is used both as MLP depth and running index; the WADL fusion is described only as "controls the balance"; and $s_i$ appears in Algorithm 2 undefined. A reader cannot reconstruct WADL from the text alone.
• Complexity analysis omits covariance inversion. The $O(\bar n d'^2)$ MHD cost silently assumes $\Sigma_a^{-1}$ is precomputed; estimation is $O(\bar n d'^2)$ and inversion $O(d'^3)$, non-negligible for Questions ($d'=301$) and Elliptic ($d'=166$), and the estimator (per-client vs. global, regularised?) is never specified.
• Narrow baseline set and no centralized bound. FedCAD (Kong et al., 2024) and FedGAD (Fang et al., 2025) are named as federated GAD methods but excluded without justification, and no centralized GAD upper bound (CoLA/GAD-NR/STIM) is reported to quantify the cost of federation.
• Statistical rigour and ranking convention. Three trials, no significance testing, and an undefined "large std → lowest rank" rule applied over only five of seven datasets — excluding Questions and Elliptic, where MV-FGAD is not universally best.
• Ambiguous terminology and data description. "Efficiency" denotes runtime, communication cost, and cross-seed variance interchangeably; the FedAvg description ("handles … non-IID") undercuts the paper's own motivation; and the Amazon vs. Amazon-all node/anomaly-rate discrepancy (10,244/6.7% vs. 11,944/9.5%) is unexplained despite prose saying only edges differ. There is also a table inconsistency (Amazon AAL+SAL AUPRC 20.84 in Table 2 vs. 19.84 in Table 5).

Reproducibility & code

Execution here was a full from-scratch reimplementation: the repository advertised in the abstract contains only a README (no training, evaluation, dataset, or ablation code), so the method and all seven baselines had to be rebuilt from the prose. The overall replication score is very low.

• Headline numbers. MV-FGAD's own AUROC reproduced within ~2–3 points on Reddit, Tolokers, Amazon, and YelpChi, but none fell inside the paper's tight ±1σ bands, and the Amazon-all AUROC reached only ~70.25 vs. the reported 90.92 — a ~20-point gap and the paper's single most important number. AUPRC matched within 1σ only on Reddit.
• Rank and margins. The best-rank claim (1.0/1.2) reproduced as ~5.0/5.2, and the Amazon-all improvement claims (+41.2%/+177.8% over classification baselines, +18.3%/+96.7% over LG-FGAD) reversed sign (roughly −4%/−25% and −3%/−23%). Part of this is because the reimplemented baselines are stronger than the paper's reported ones, so the rank collapse is partly a replication-faithfulness limitation rather than a demonstrated paper error.
• Efficiency. The runtime advantage inverted (MV-FGAD slowest on Questions, ~8x slower than the reimplemented FedCLGN vs. the claimed 0.33%), the FedCLGN Elliptic-OOM did not reproduce (FedCLGN completed; MV-FGAD peaked at ~60 GB, the most of any method), and the round-10 communication advantage inverted (MV-FGAD lowest at round 10, degrading by round 100). The runtime/OOM inversions depend on the original heavy baselines being unavailable and on the larger test GPU, so they are not fully attributable to the paper — but they mean the efficiency thesis is unconfirmed.
• Numerical instability. The method as written required a critical undocumented fix (the symmetric normalization around Eq. 6/7 is undefined for the negative similarities the method produces, giving NaN on every YelpChi client) plus major stabilizations of the relation-matrix and propagation steps and a fix for FedAvg being poisoned by a single NaN client update. None of these are in the paper or the (empty) repo.
• Partitioning and coverage. Naive Louvain produced only 6 of the required 10 clients on Amazon-all, needing an undocumented coercion rule. The multi-view, scoring, and aggregation ablations (Tables 5–7), the client-number sweep (Fig. 9), and the score-distribution case studies (Figs. 5/10/14/15) could not be attempted at all, as no code isolates single views, swaps the distance metric, or renders the density plots.
• Parameter size. MV-FGAD's 0.022 MB matched exactly, but the three baselines collapsed to one shared 0.054 MB value (they share one reimplemented backbone), so the "comparable to FedCLGN, <1/5 of FGAD/LG-FGAD" contrast could not be reconstructed.

Recommended Changes

Essential

• Release runnable code. Populate the advertised repository with the training entry point, the view/scoring/aggregation ablation toggles, dataset preprocessing (including the Amazon vs. Amazon-all node treatment), the runtime/memory harness behind Figs. 2 and 6, the score-distribution plotting code, and pinned dependencies — this is the precondition for any of the effectiveness/efficiency claims to be checkable.
• Reconcile the Amazon-all result. Release the exact config that yields AUROC 90.92 and explain why it exceeds Amazon by ~13 points despite sharing the same underlying data; if the number cannot be reproduced from released code, correct it. Report per-seed values.
• Specify the stabilizations the method needs to run. State exactly how the Eq. (6)/(7) normalization handles non-positive similarities (ReLU/clamp?) and any L2-normalization applied in the WADL relation matrix and propagation, so the method does not diverge to NaN.
• Write out the WADL recursion. Give explicit equations for $\tilde{\mathbf{S}}^{(l)}$, $\mathbf{H}_c^{(l)}$, $\mathbf{H}_d^{(l)}$, $\mathbf{H}_w^{(l+1)}$ and the $\beta$-fusion, and add a quantitative argument (or reframe as an empirically-tested hypothesis) for the "weak signals no longer weak" mechanism.
• Bring prose in line with the tables. Soften (i) "combining any two views consistently improves" (AAL+WADL regression on Amazon), (ii) "Across all datasets, MHD…" (Questions/Elliptic exceptions), and (iii) "multi-view … across all datasets" (single-WADL wins on Elliptic).
• Document the missing hyperparameters and protocol. State the margin $m$, per-view MLP depths/activations/dropout, the covariance estimator for $\Sigma_a$ (per-client vs. global, regularisation), the Louvain resolution/seed and the client-count coercion rule, the three seeds, and resolve the "200 epochs / 100 rounds" ambiguity.
• Correct the MHD complexity. Add the $O(\bar n d'^2 + d'^3)$ cost of estimating and inverting $\Sigma_a$, and acknowledge the quadratic $O(\bar n^2 d')$ relation-matrix term when discussing efficiency and large-graph memory.

Suggested

• Fix the runtime/efficiency claims against faithful baselines. Re-measure runtime, memory, and parameter size against the original public implementations of FedCLGN/FGAD/LG-FGAD, and anchor ratios like "0.33%" to the specific dataset/baseline rather than the framework as a whole.
• Justify or expand the baseline set. Add FedCAD (Kong et al., 2024) and FedGAD (Fang et al., 2025) or state per method why they were excluded.
• Add a centralized upper bound. Include one centralized GAD method (CoLA/GAD-NR/STIM) per dataset to quantify the FL-vs-centralized gap.
• Formalise the ranking rule and add rigour. Define "large std" quantitatively, report ranks with and without the penalty over all seven datasets, raise the trial count, and add paired significance tests on the headline cells.
• Tighten terminology and descriptions. Reserve "efficiency" for runtime/communication and use "robustness"/"reliability" for variance; soften the FedAvg non-IID description; reconcile the Table 2 vs. Table 5 Amazon AUPRC entry; and add the Amazon vs. Amazon-all node-selection rule.
• Rephrase overclaimed qualitative language. State the case-study conclusions in terms of overlap area / distributional distance rather than peak height, and use "generally reduces/sharpens" instead of "maximizes/minimizes across client subgraphs."