{"id":"65758e83-4ef5-4910-812a-7bc88afc4622","arxiv_id":"2502.05472","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"DSGC improves K-way signed graph clustering by denoising edge signs, augmenting graph structure, and training a weak-balance encoder that separates negatively linked nodes.","lead":"This paper presents DSGC, a method for clustering signed networks, which are graphs with friend and enemy edges. It first repairs and adds edges, then learns node representations that keep negatively linked nodes apart, achieving stronger clustering results than prior methods.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VS-R as specified (Eq. 5, δ±=±1) cannot flip any edge on the sparse SSBM graphs evaluated: Γ is dominated by the edge's observed sign (±1) because higher-order walks contribute ~0.01–0.1, so the flip thresholds never fire; the large VS-R gains in Table 2/Fig.","rationale":"The paper's central claim is that DSGC's two rewiring modules plus weak-balance encoder make it consistently more accurate and robust than prior signed clustering methods, establishing a new benchmark. The load-bearing sub-claim is that VS-R denoises by correcting noisy edges via the non-noise score Γ_ij (Eq. 3) and thresholds δ± (Eq. 5). My walk-count analysis in the evaluated SSBM regime (N=1000, p=0.01–0.04, K=4–10) shows the mechanism cannot fire: for any existing edge, the l=1 term of Γ equals the edge's observed sign (±1), all higher-order walk terms combined have magnitude ≈0.01–0.1, and the thresholds δ±=±1 require corrections exceeding 2 in magnitude to trigger a flip. Hence Â=A and VS-R is a no-op, yet Table 2, Fig. 4, and Fig. 13(a) report substantial VS-R effects and a peaked sensitivity curve. The only consistent reading is that the experiments used an undocumented rule (e.g., excluding the l=1 term, near-zero thresholds, or a sign-mismatch override), meaning the paper as written does not describe the system it evaluates. This directly undermines the robustness mechanism advertised in the abstract; it is not merely a matter of missing error bars or under-specification. This sharpens the reader's weakest_assumption, which worried that Γ/δ might misclassify edges in some regime; the sharper problem is that Γ/δ as reported cannot classify anything in the tested regime. The released code makes a decisive check cheap. If the check confirms inertness, the manuscript should be rejected in its current form (the contribution and all VS-R attributions would need re-derivation and re-validation); if the check refutes my walk-count analysis (e.g., the code implements a working variant), the reader's CONDITIONAL verdict stands. I therefore move the verdict from CONDITIONAL to REJECT, conditioned on the outcome of the code check.","tokens_in":25127,"tokens_out":27517,"duration_ms":258738,"concrete_test":"Instrument the released code (github.com/yaoyaohuanghuang/DSGC) and run VS-R on SSBM(1000,5,0.01,0.04) with L′=3, δ+=1, δ−=−1; count how many entries of  differ from A and record the Γ distribution over flipped vs. retained edges. If zero (or <0.1%) edges change, the described mechanism is inert in its own evaluation regime, so Table 2 and Fig. 4 cannot be produced by the method as written, and the paper must disclose the actual VS-R rule and re-run ablations under the stated hyperparameters. If many edges change, report the implemented thresholds/rule, since it does not match Eq. (5)–App. D.","verdict_should_be":"REJECT","load_bearing_attack":"In VS-R (§4.1.1), Γ_ij (Eq. 3) includes the l=1 term α_1(μ+_1 − μ−_1), which for any existing edge equals exactly its observed sign ±1; the higher-order terms Σ_{l≥2} α_l(μ+_l − μ−_l) have expected magnitude ~0.02–0.1 on the SSBM graphs in Tables 1–2 (e.g., μ−_2 ≈ 2·(N/K)p² ≈ 0.04 at N=1000, K=5, p=0.01). With the reported thresholds δ+=1, δ−=−1 (App. D; note Eq. 5 requires δ−<0 but App. D says both are '1'), the flip branches of Eq. (5) require Γ>1 or Γ<−1, i.e., higher-order evidence must overcome the edge's own sign by more than 2 in magnitude. That cannot happen at p=0.01–0.04 (the expected corrections are ~0.04, and even multi-sigma excursions stay below ~0.2), so Â=A exactly and VS-R is inert. Yet Table 2 reports VS-R improving spectral baselines by up to 61 ACC points and halving the violation ratio, and Fig. 4 shows w/o-VS-R differing from DSGC; these outcomes are impossible with Eq. (5) as written. They require an unstated rule (e.g., omitting the l=1 term from Γ, δ≈0, or a sign-mismatch override). The abstract's robustness claim rests on this denoising mechanism, so the reported experiments do not validate the described method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DSGC, an unsupervised deep signed graph clustering framework built on Weak Balance Theory. It contains two graph rewiring preprocessing modules: Violation Sign-Refine (VS-R), intended to correct noisy edge signs using high-order walk counts, and Density-based Augmentation (DA), which adds positive edges within clusters and negative edges across clusters. The framework then employs a two-channel signed graph encoder that aggregates positive and negative walks while deliberately omitting the 'enemy of my enemy is my friend' rule, and optimizes a differentiable soft-assignment clustering loss. The authors evaluate on 20 synthetic SSBM configurations spanning noise level, sparsity, graph size, and cluster count, plus two real-world graphs, and report consistent superiority over 17 baselines in ACC, NMI, ARI, and F1.","tokens_in":25600,"tokens_out":10366,"duration_ms":98700,"significance":"The framework is clearly motivated and the experimental scope is broad: 20 SSBM configurations, two real-world graphs, a public code release, and ablations of each component. The weak-balance encoder and the soft violation-minimization loss are coherent and reasonable design choices. However, the central robustness claim rests on the VS-R denoising mechanism, and that mechanism, as specified in Eq. (5) with the thresholds reported in App. D, is internally inconsistent and cannot have produced the reported ablation and preprocessing gains. Until this is resolved, the stated contribution is not supported by the experiments as described.","major_comments":[{"comment":"The VS-R rule cannot flip any edge on the evaluated SSBM graphs. For any existing edge, the l=1 term in Eq. (3) equals the observed sign A_ij ∈ {±1} exactly, so Γ_ij = ±1 + Σ_{l≥2} α_l (μ+_l − μ−_l). On the p=0.01–0.04 SSBM graphs used in Tables 1–2, the l≥2 corrections have expected magnitude below roughly 0.3 even at p=0.04, and closer to 0.02–0.05 at p=0.01–0.02 (e.g., (A+)^2_ij ≈ (N/K)p^2 ≈ 0.02 at N=1000, K=5, p=0.01). Consequently, for a negative edge Γ_ij remains above −1, and for a positive edge Γ_ij sits near +1 without crossing the δ+=1 threshold in a way that would change the sign. With δ+=1 and δ−=−1, neither the 'set to +1' branch nor the 'set to −1' branch of Eq. (5) can be triggered for an existing edge; hence  = A identically. This contradicts the large ACC/NMI improvements and violation-ratio reductions attributed to VS-R in Table 2 and the ablation curves in Fig. 4. The authors must either exclude the l=1 term from Γ, choose thresholds near 0, or specify a different correction rule, and then present experiments that actually exercise that rule.","section":"§4.1.1, Eqs. (3)–(5) and App. D"},{"comment":"The reported violation-ratio reductions are impossible if VS-R is the identity. If no edge signs are changed, the violation ratio (defined as the ratio of violated to non-violated edges) is unchanged; however, the text states that VS-R 'consistently reduces the violation ratio' across the tested configurations. The discrepancy between the described algorithm and the reported Table 2 numbers must be resolved before the robustness claim can be evaluated.","section":"Table 2 and §5.4"},{"comment":"The instruction that 'optimal performance is achieved when both δ+ and δ− are set to 1' directly violates the constraint δ− < 0 stated after Eq. (5), and it makes the 'retain' interval δ− ≤ Γ ≤ δ+ degenerate (it becomes Γ = 1 exactly). Fig. 13(a) shows ACC against a single δ axis up to 12, but the text does not specify whether δ+ and δ− are varied together, separately, or with what sign convention. This is not a cosmetic issue: the threshold values completely determine whether VS-R modifies any edge, so the reported hyperparameter analysis must be made consistent with the algorithm definition.","section":"App. D and Eq. (5)"}],"minor_comments":[{"comment":"Specify whether L+ and A− in the clustering loss are built from the original adjacency matrix A or from the augmented A′′, since the encoder uses the latter; the loss–encoder consistency is otherwise unclear.","section":"Eq. (14)"},{"comment":"The paper says '15 baselines', but App. B enumerates 9 spectral, 2 deep signed (SiNE, SNEA), and 6 deep unsigned methods, totaling 17; the count should be aligned.","section":"Abstract and §5.2"},{"comment":"All results are averages over 5 runs without standard deviations or significance tests; given the claim of consistent superiority, at least standard deviations or error bars should be reported.","section":"§5.2 and App. G"},{"comment":"The symbol L is overloaded: it denotes both the signed Laplacian baseline (Table 1) and the clustering loss (Eq. (14)); rename one to avoid confusion.","section":"Notation"},{"comment":"The phrase 'first Deep Signed Graph Clustering framework' should be softened, as SSSNET [16] is a deep signed clustering method, albeit semi-supervised.","section":"§1"},{"comment":"A one-line proof or citation of the walk-count identity in Eq. (2) would help readers verify the lemma.","section":"Lemma 1"}],"recommendation":"major_revision","confidential_remarks":"The central problem is the mismatch between the specified VS-R rule and the reported experiments. If the authors respond with a corrected algorithm and rerun the relevant experiments, the paper could be acceptable; if the reported gains came from a substantially different rule, Section 4.1.1 and Table 2 would need a full rewrite. I would also ask the editor to verify whether the 'first deep signed graph clustering' claim is defensible given SSSNET. The rest of the framework, especially the DA and encoder components, appears plausible and worth a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the stress-test note holds up, and it is the main thing you should know about this paper. In Eq. (3), Γ_ij includes the l=1 term α_1(μ+−μ−), which for an existing edge is exactly its observed sign. With δ+=1 and δ−=−1 (or absolute values of 1, as App. D loosely says), flipping a positive edge to negative requires the higher-order walk evidence to push Γ below −1, i.e., the higher-order correction must be less than −2. On the sparse SSBM graphs (p ≈ 0.01–0.05), those corrections are tiny — expected magnitudes around 0.02–0.1, with maximum possible contributions bounded well below 2. So the flip branches of Eq. (5) essentially never fire. VS-R is inert on the very graphs where Table 2 credits it with enormous improvements. Either the implementation differs from the text (e.g., omitting the l=1 term or using thresholds near zero), in which case the paper misdescribes its method, or the reported VS-R gains are not reproducible. This is a load-bearing inconsistency, not a cosmetic one, because the robustness claim rests on VS-R.\n\nThe rest of the paper is more defensible. The combination of density-based augmentation, a weak-balance encoder, and a differentiable signed cut loss is a sensible pipeline, and the K>2 weak-balance motivation for dropping 'EEF' is legitimate. The experimental sweep is broad: 20 SSBM configurations, four metrics, ablation studies, real-world visualizations. The ablation suggests the components matter, though without the VS-R inconsistency resolved, those results are hard to trust.\n\nOther soft spots are real but secondary. The 'first deep signed graph clustering framework' claim needs qualification relative to SSSNET (which is semi-supervised); Lemma 1 is stated without proof or reference; no error bars or significance tests; baseline hyperparameters are not given. These are easy to fix.\n\nWho is this for? People working on signed graph clustering and robustness in GNNs will find the weak-balance encoder idea worth thinking about. But the paper, as it stands, should not be accepted as a reliable benchmark until the authors clarify exactly what VS-R does and release code to confirm it. A serious referee should engage with it, because the underlying idea is plausible and the experiments are extensive — but the revision needs to resolve the VS-R contradiction. I would not cite the method in its current form.","headline":"The weak-balance pipeline is interesting, but VS-R as written cannot flip any edge on the sparse graphs tested, so the paper's central robustness claim is contradicted by its own equations.","tokens_in":740,"tokens_out":757,"would_cite":false,"duration_ms":48208,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that replacing Social Balance's 'enemy of my enemy is my friend' with Weak Balance's relaxed rule, plus edge-sign rewiring, makes deep K-way signed clustering more accurate and robust than every baseline it tests.","keywords":["signed graph clustering","weak balance theory","social balance theory","graph neural networks","graph rewiring","noise robustness","K-way clustering","signed stochastic block model"],"falsifier":"Generate SSBM(1000, 5, 0.01, $\\eta$) graphs at $\\eta = 0.15$, $0.25$, and $0.35$ and measure DSGC's ACC against SPONGE: if the gap shrinks or reverses as $\\eta$ approaches $0.4$, the robustness claim is confined to the low-noise regime the paper tested ($\\eta \\le 0.08$). A second check is to cluster a graph with a known K=2 structure inside a K=5 layout and see whether VS-R, which assumes violations are noise, erases legitimate boundaries.","tokens_in":24977,"feed_emoji":"⚖️","tokens_out":16226,"duration_ms":138160,"temperature":0.7,"pith_summary":"Signed graphs mark every link as positive or negative—friend or enemy, trust or distrust—which makes them expressive models for social and biological networks, but also fragile targets for clustering. The paper argues that the classical Social Balance rule 'an enemy of my enemy is my friend' implicitly presumes only two clusters and, in graphs with K>2 communities, pulls boundary nodes toward the wrong cluster. It proposes DSGC, an unsupervised deep clustering framework built on Weak Balance Theory, which replaces that rule with the weaker 'an enemy of my enemy might be my enemy.' Before learning, DSGC rewires the graph: Violation Sign-Refine flips the signs of edges it judges noisy, and Density-based Augmentation adds positive edges inside clusters and negative edges across clusters. Across the paper's 20 SSBM configurations and two real-world networks, DSGC reports higher accuracy and robustness than every spectral and deep baseline it compares against, with ablations tracing most of the gain to the rewiring steps.","feed_headline":"Weak balance beats balance for K-way signed clustering","feed_subtitle":"Replacing strict balance with weak balance widens cluster boundaries and withstands noisy edge signs.","key_machinery":"The load-bearing machinery is the rewiring pipeline defined by Eqs. (2)–(7). Definition 1 redefines a positive walk as one whose edges are all positive, and a negative walk as one with exactly one negative edge and the rest positive; Lemma 1 then expresses the difference of their counts as a matrix identity involving powers of the positive adjacency and one negative adjacency. The non-noise score $\\Gamma_{ij}(L') = \\sum_{l=1}^{L'} \\alpha_l (\\mu^+_l(i,j) - \\mu^-_l(i,j))$ aggregates these counts up to length $L'$ with weights $\\alpha_l$ that decay with $l$, and Eq. (5) thresholds $\\Gamma$ at $\\delta_+$ and $\\delta_-$ to decide whether each edge should become $+1$, stay as is, or become $-1$. Density-based Augmentation follows, adding edges along long positive and negative walks via the matrix powers in Eq. (6). On the encoding side, the signed convolutions of Eqs. (9)–(10) are the second piece of machinery: the positive branch aggregates along positive walks, and the negative branch uses $-\\bar{A}^-$ so that negative edges repel embeddings, widening cluster boundaries. The soft-assignment matrix $\\Pi$ and the differentiable clustering loss of Eq. (14) with the volume regularizer of Eq. (15) close the loop, making the whole system trainable end-to-end.","core_discovery":"The central claim is that K-way signed graph clustering should be built on Weak Balance Theory rather than Social Balance Theory, and that the difference is measurable: the 'friend-of-my-enemy' reasoning of strict balance narrows cluster boundaries, while the relaxed 'enemy of my enemy might be my enemy' keeps nodes connected by negative edges well separated. To make that work, DSGC operates in two stages. In the preprocessing stage it overwrites the adjacency matrix: Violation Sign-Refine (VS-R) computes a non-noise score $\\Gamma_{ij}$ from redefined positive and negative walk counts and flips edge signs that fall beyond thresholds $\\delta_+$ and $\\delta_-$; Density-based Augmentation (DA) then takes powers of the refined adjacency matrices to add positive edges within clusters and negative edges across clusters. In the learning stage, a signed convolutional encoder produces separate positive and negative embeddings, with the negative branch carrying an explicit minus sign so that negatively linked nodes push apart, and a regularized soft-assignment loss drives the final K-way partition. The evidence for the claim is a 20-setting SSBM benchmark plus the S&P1500 and Rainfall networks, where DSGC beats every baseline on ACC, NMI, ARI, and F1; ablations show that DA contributes the most, VS-R second, and that dropping the 'EEF' rule matters less than keeping the negative repulsion term in the encoder.","pith_inferences":["The robustness claim is probably confined to a moderate-noise window: the $\\Gamma$ score uses hand-set weights and thresholds, so when sign flips become frequent ($\\eta$ beyond roughly 0.15–0.2) the high-order walk 'vote' itself becomes unreliable and DSGC's advantage over spectral methods should erode.","The augmentation via matrix powers is a form of graph diffusion; replacing the hand-picked powers $m_+$ and $m_-$ with learnable or standard diffusion kernels (e.g., personalized PageRank) could yield similar or better rewiring with fewer hyperparameters.","The ablation showing DA as the most influential component suggests that edge addition, not sign correction, drives most of the performance; a direct test would be a version of DSGC that only runs DA on the raw graph, skipping VS-R entirely on low-noise graphs.","The argument that 'EEF' narrows boundaries is qualitative; converting the SoEN metric into a formal notion of cluster-boundary width could make the claim quantitative and testable across different balance axioms."],"forward_implications":["Table 2 shows that applying VS-R alone improves the ACC and NMI of spectral baselines such as BNC, SPONGE, and SPONGE_sym, so the denoising step can be used as a general preprocessor for other signed clustering methods.","If the EEF rule is as harmful as the paper argues, signed GNNs built for link prediction should not be transplanted unchanged into clustering; clustering-oriented encoders should derive their aggregation from Weak Balance instead.","The differentiable reformulation of the signed clustering objective means the NP-hard partition problem can be optimized end-to-end without labels, which is what allows the whole pipeline to train jointly.","The reported gains across K = 4 to 10 and N up to 1200 indicate that, within the tested regime, the method scales in cluster count and graph size while retaining an advantage over spectral baselines."],"supporting_citations":[{"why":"Defines Balanced Normalized Cut for K-way signed clustering; it is the spectral baseline DSGC must beat and the cut criterion its loss reformulates.","marker":"[5]"},{"why":"Introduces SPONGE, the strongest spectral baseline, and supplies the SSBM synthetic setup and the real-world datasets (S&P1500, Rainfall) used for evaluation.","marker":"[6]"},{"why":"States Weak Balance Theory, the relaxed principle ('an enemy of my enemy might be my enemy') that DSGC builds its preprocessing and encoder around.","marker":"[7]"},{"why":"Presents signed graph convolutional networks whose balance-theory aggregation DSGC adapts and modifies for clustering.","marker":"[9]"},{"why":"Defines positive/negative walks via parity of negative edges; DSGC redefines these walks for K-way clustering in Definition 1.","marker":"[10]"},{"why":"Formulates Social Balance Theory, the 2-way theory whose 'EEF' principle DSGC argues narrows cluster boundaries.","marker":"[15]"},{"why":"The prior deep signed clustering network (semi-supervised) that DSGC contrasts with its fully unsupervised setting.","marker":"[16]"},{"why":"Establishes spectral analysis of signed graphs for clustering and links balance theory to 2-way partition; it is the basis of the signed Laplacian baselines.","marker":"[22]"}],"fun_headline_variants":["Weak balance theory cracks noisy signed clustering","Signed graph clustering that shrugs off noisy edges","Relaxing balance theory improves signed graph clustering","Why weak balance wins for robust signed clustering"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method stands or falls on the assumption that noisy edges are sparse enough for a hand-tuned score over high-order walks, with fixed weights and two thresholds, to tell real edges from noisy ones before learning begins; if that rewiring mislabels edges, every later stage trains on a corrupted graph.","fun_headline_variants_meta":{"raw":{"variants":["Weak balance theory cracks noisy signed clustering","Signed graph clustering that shrugs off noisy edges","Relaxing balance theory improves signed graph clustering","Why weak balance wins for robust signed clustering"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000621,"raw_usage":{"total_tokens":2939,"prompt_tokens":1067,"completion_tokens":1872,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":683,"completion_tokens_details":{"reasoning_tokens":1825}},"tokens_in":683,"tokens_out":1872,"duration_ms":17021,"temperature":1.0,"reasoning_tokens":1825,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T19:11:08.438026+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate SSBM(1000, 5, 0.01, $\\eta$) graphs at $\\eta = 0.15$, $0.25$, and $0.35$ and measure DSGC's ACC against SPONGE: if the gap shrinks or reverses as $\\eta$ approaches $0.4$, the robustness claim is confined to the low-noise regime the paper tested ($\\eta \\le 0.08$). A second check is to cluster a graph with a known K=2 structure inside a K=5 layout and see whether VS-R, which assumes violations are noise, erases legitimate boundaries.","supporting_citations":[],"review_version":1}