{"id":"d6b87b4d-ebf0-4fb3-99fc-4ded19c8cbf7","arxiv_id":"2501.05590","paper_version":4,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Negative votes on Menéame expose ideologically extreme users, such as pro-Russia accounts, that positive-only embeddings fail to distinguish from ordinary left-wing users.","lead":"Using voting data from the Spanish news platform Menéame, the authors add negative interactions (downvotes) to a signed network and show that this reveals highly polarized, antagonistic users that positive-only analysis misses. A generalist should read it because it offers a practical way to detect extreme and potentially foreign-influenced accounts in online debates.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'only' in the central claim is not established: a single unsigned baseline (CA) is compared, and CA's degree sensitivity could produce the tail differences without negative ties; a positive-only spectral control is needed.","rationale":"The paper is a solid empirical contribution: the dataset and code are public, outlet ideology is validated against Twitter and PoliticalWatch, and the Appendix A.9 null model is a genuine control for the algorithm-choice objection. The finding that SHEEP separates RT-upvoting users from the general left-wing faction is plausible and potentially important. My stress-test targets the strongest formulation of the claim, the word 'only' in the abstract. The design compares exactly one signed method against exactly one unsigned method, and the Appendix A.9 null model does not fill that gap: it changes the signed input by adding artificial negative links, but it never asks whether an unsigned method with SHEEP-like regularization would already recover the RT extremes. The paper's own degree-correlation statistics show that SHEEP and CA differ sharply in how they treat low-degree users, so the tail differences could be a regularization effect rather than a negative-tie effect. A positive-only spectral embedding with the same normalization is the minimal control that isolates the contribution of negative ties. If that control separates RT-upvoters, the central 'only' claim fails; if it does not, the claim is considerably strengthened. This concern does not warrant rejection; conditional acceptance tied to running this control is appropriate. The reader identified the missing-edge coding assumption, which is real, but I see the single-method unsigned baseline as the more load-bearing gap for the 'only' claim, hence partial agreement.","tokens_in":30300,"tokens_out":8405,"duration_ms":87978,"concrete_test":"Run SHEEP with the same normalization, PCA projection, and degree filters on the unsigned positive-only network, i.e., set A^- = 0 in Eq. (1) so the repelling Laplacian reduces to L = D^+ - A^+, for both Russia-Ukraine and Broad Politics. Also run one additional unsigned embedding (e.g., normalized Laplacian eigenmaps or node2vec) as robustness. Then check whether users who positively voted for RT (red circles in Fig. 3C) appear in the tails of these unsigned embeddings, and whether the tail-vs-degree relationship matches SHEEP or CA. If any unsigned baseline places RT-upvoters at the extreme, the 'only negative ties' claim is falsified; if none does and the SHEEP tail persists after matching users on degree, the claim is substantially strengthened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim ('only by incorporating negative ties can we identify ideologically extreme users') is load-bearing and currently rests on a single comparison: SHEEP on the signed network versus CA on the unsigned network. This comparison confounds the data source (negative ties) with the algorithm. CA uses chi-square residual standardization, which places low-degree users at the extremes, whereas SHEEP's degree-normalized repelling Laplacian pulls low-degree users toward the center; the paper itself reports degree-embedding correlations of 36% for SHEEP versus 22% for CA in the Russia-Ukraine topic (Section 3.2.1). Appendix A.9 replaces real negative ties with artificial links of weight -50 and shows that this CA-like treatment makes SHEEP correlate with CA, but it does not test whether a positive-only version of the same spectral algorithm, or any other unsigned method, already separates RT-upvoting users from left-wing users. Because the abstract asserts uniqueness, the absence of such a control leaves open that the hidden extremes are an artifact of SHEEP's regularization rather than of negative-tie information. The reader's coding concern about zero-weight missing edges is related, but the more decisive gap is the missing baseline comparison for the word 'only'.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies polarization on the Menéame social news platform using a dual-method approach: SHEEP, a spectral embedding for signed networks, and Correspondence Analysis (CA) on unsigned networks. The authors construct user-to-user and user-to-outlet networks from voting data, extract topics with BERTopic and hSBM, and compare how the two embedding methods position users and news outlets on political dimensions. The core empirical claim is that only the signed method (SHEEP) identifies ideologically extreme users who engage in antagonistic behavior, such as users who upvote Russia-affiliated outlets in the Russia-Ukraine topic, while the unsigned method (CA) conflates these users with ordinary left-wing users. The paper also contributes a publicly available dataset and code, and validates outlet ideology against external Twitter-based and PoliticalWatch benchmarks.","tokens_in":30549,"tokens_out":3069,"duration_ms":32315,"significance":"If the central claim is established, the paper would provide a substantive empirical demonstration that negative ties carry information about hidden ideological extremes that positive-only network methods miss. The dataset release and the external validation of outlet embeddings are clear strengths: the Twitter and PoliticalWatch comparisons (Section 2.6, Figures A8-A9) give independent anchors for interpretation, and the analysis is generally careful and transparent about parameter choices. The claim of uniqueness, however, is currently supported only by a single comparison of two different algorithms on two different network representations, which makes the significance conditional on additional controls.","major_comments":[{"comment":"The central claim that 'only by incorporating negative ties can we identify ideologically extreme users' is not established because the comparison confounds the data source (signed vs. unsigned) with the algorithm (SHEEP vs. CA). The paper compares SHEEP on the signed network to CA on the unsigned network, but CA uses chi-square residual standardization that places low-degree users at the extremes, whereas SHEEP's degree-normalized repelling Laplacian pulls low-degree users toward the center; the paper itself reports degree-embedding correlations of 36% for SHEEP versus 22% for CA (Section 3.2.1). The null model in Appendix A.9 replaces missing edges with artificial negative ties of weight -50 and shows that this CA-like input makes SHEEP correlate with CA, but it does not test whether a positive-only version of the same spectral algorithm, or any other unsigned spectral method, already separates RT-upvoting users from left-wing users. A positive-only SHEEP control, or an equivalent unsigned spectral embedding with the same degree normalization, is needed to rule out that the observed tail differences are an artifact of algorithmic regularization rather than of negative-tie information.","section":"Abstract; Section 3.2.1; Appendix A.9"},{"comment":"The coding of missing interactions is load-bearing: the signed adjacency matrix sets A_ij = 0 both when users have not interacted and when their positive and negative interactions balance, while the unsigned network treats missing edges as absent and the null model (Appendix A.9) treats them as negative ties of weight -50. The central finding that negative ties reveal hidden extremes depends on zero-weight missing edges being the correct coding. If missing edges in fact signal dissimilarity, the extremes might be recoverable without explicit negative ties. The authors should provide a robustness analysis that varies the coding of non-interactions, for example by assigning them small negative weights or by thresholding low-weight edges, and show that the RT-upvoting users remain extreme under these alternatives.","section":"Section 2.3, footnote 4"},{"comment":"The null model that is used to argue that differences between SHEEP and CA are due to negative ties relies on an arbitrary artificial weight of -50 for all missing edges, chosen only because it is 'a similar magnitude to the largest positive links'. The reported correlations between the SHEEP null model and CA (0.83 for Russia, 0.94 for Politics) may depend on this specific value. The authors should perform a sensitivity analysis over a plausible range of negative weights and report whether the conclusion that the null model aligns with CA is stable; without this, the null-model comparison does not tightly constrain the role of real negative ties.","section":"Appendix A.9"},{"comment":"The authors remove stories that only received positive votes before constructing the networks, stating that such stories 'do not polarize the discussion'. This is a post-hoc filter that could affect which users appear extreme, because it removes a potentially large amount of positive interaction data and changes degree distributions. The robustness of the RT-upvoting-extreme finding to this filtering decision should be checked, for example by repeating the analysis with all stories included or with an alternative criterion for controversial stories, and verifying that the main tail differences between SHEEP and CA persist.","section":"Section 2.3"}],"minor_comments":[{"comment":"The footnote states that neutral interactions (zero-sum votes) constitute only 0.74% of the dataset, but it then equates non-interaction with a null sum. This is a strong assumption that deserves a fuller justification, especially because the unsigned network and the null model treat non-interaction differently.","section":"Section 2.3, footnote 4"},{"comment":"The caption says 'schematic representation of one of the stores in the platform', which appears to be a typo for 'stories'.","section":"Figure 1 caption"},{"comment":"The sentence 'far-left extremists use negative votes strategically to target the opposing extreme faction' uses causal and intentional language; the analysis demonstrates a correlation between extreme structural position and negative voting behavior, not strategic intent. A more neutral phrasing would be appropriate.","section":"Section 3.2.2"},{"comment":"The binning procedure divides the number of users by a normalization constant set to 1.5, but no sensitivity analysis is given for this choice; since the k-means clusters and the resulting interpretation of faction behavior depend on the bins, a brief robustness note would be useful.","section":"Appendix A.6"},{"comment":"The word 'only' in the abstract and introduction overstates the current evidence, which is based on one platform, two topics, and one pair of methods. I suggest softening the claim to reflect the scope of the empirical comparison, for example 'in this dataset' or 'with the methods considered here'.","section":"Abstract; Section 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution with valuable data release and external validation, and the central question is timely. However, the headline claim of uniqueness is currently supported by a single method comparison that conflates algorithm with data source, and the null-model and missing-edge coding assumptions need robustness checks. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection. The self-citation to SHEEP is not circular because SHEEP is independently published and validated, but the authors should be careful in revision to frame the 'only' claim conditionally on the methods compared."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe useful thing here is the new public Menéame dataset and the clean demonstration that, on this platform, negative ties separate a specific group of extreme users—people who upvote RT and other pro-Russia outlets—from the broader anti-NATO/left-wing faction that a positive-only Correspondence Analysis lumps them with. The external validation is real: RT voting, the Twitter-derived outlet ideology, and the PoliticalWatch cross-check all line up. The paper is also honest about its methods. The null-model exercise in A.9, which feeds SHEEP a network with all missing edges set to -50, is a genuine attempt to show that the difference comes from the negative-tie information rather than from SHEEP's algorithm.\n\nThe soft spot is the 'only.' The abstract says only by incorporating negative ties can we identify these users, but the support for that uniqueness claim is one comparison: SHEEP on the signed network versus CA on the unsigned network. These differ in more than the sign of the edges. CA's standardization pushes low-degree users to the extremes; SHEEP's degree-normalized Laplacian pulls them toward the center. The paper's own degree-embedding correlations (36% vs 22%) show exactly this. Without a positive-only version of the same spectral algorithm, or at least a second unsigned method, the tail differences could be an artifact of SHEEP's regularization rather than of the negative-tie information. The A.9 null model does not close that gap: it never removes the real negative edges and re-runs SHEEP on the positive-only network. So the central empirical finding—negative ties help in practice—is solid, but the claim that no unsigned method could do it is not yet established. The treatment of missing edges as zero weight (footnote 4) is a related assumption; the -50 null-model weight is arbitrary, and correlations are reported without error bars. These are moderate limitations, not fatal ones.\n\nWho benefits: anyone working on signed networks and polarization, and platform-governance people who care about identifying state-linked or adversarial user groups from interaction data. The dataset is reusable. The overclaim is fixable with one additional baseline experiment.\n\nMy recommendation: send it out. A serious referee can push for the missing control and a sensitivity analysis on the null-model weight. Accept after revision, not in current form.","headline":"New Menéame dataset and a careful signed-vs-unsigned comparison, but the 'only' claim about negative ties needs an additional unsigned baseline to hold.","tokens_in":31081,"tokens_out":2147,"would_cite":true,"duration_ms":21170,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["89.65.-s"],"model":"deepseek-v4-flash","headline":"The paper argues that on the Spanish news-aggregator Menéame, only signed networks that include downvotes can separate genuinely extreme users—such as pro-Russia voters in the Ukraine war—from ordinary left-wing critics, because…","keywords":["signed networks","negative ties","structural polarization","Menéame","SHEEP embedding","correspondence analysis","social media voting","online extremism"],"falsifier":"Re-run the Russia-Ukraine user network with one change: give the zero-weight missing edges a small negative weight calibrated to the observed downvote base rate on Menéame rather than treating them as neutral, and re-compute the SHEEP extremes. If the RT-upvoting clique then stops being uniquely extreme, or if CA recovers the same tails under that coding, the paper's central distinction collapses. A complementary check: among user pairs with no recorded interaction, look at the sign of the first vote they later exchange; a predominantly negative sign would mean absence itself carries a whiff of dissent, contradicting the neutrality assumption the result depends on.","tokens_in":30076,"feed_emoji":"👎","tokens_out":9996,"duration_ms":89747,"temperature":0.7,"pith_summary":"This paper asks what negative interactions—downvotes on comments and stories—reveal about online political polarization that positive-only data cannot. Using eight months of voting data from Menéame, a Spanish news-aggregator platform, the authors build two networks of the same users, one signed and one positive-only, and embed both with two spectral methods. The resulting maps largely agree on who sits on the left and right, but they diverge at the extremes: only the signed method separates users who upvote pro-Russian outlets such as RT from general anti-NATO left-wing users, and only it tags far-left users who systematically downvote the far right. Relatively few votes on the platform are negative (about 3%), yet the paper argues they carry the information that marks the most antagonistic users—the ones positive-only methods place in the same bucket as their less confrontational ideological peers. If correct, the implication is that polarization measures built on likes or upvotes alone capture the ideological center while mis-ranking the fighting edge.","feed_headline":"Downvotes unmask extremists that positive votes cannot find","feed_subtitle":"A signed-network method separates pro-Russia users from ordinary left-wing critics on Menéame; positive-only maps cannot.","key_machinery":"The central object is SHEEP (Signed Hamiltonian Eigenvector Embedding for Proximity), a spectral embedding built on the repelling Laplacian $L_r = D^+ - A^+ - D^- - A^-$: positive edges act as springs that pull connected nodes together while negative edges act as anti-springs that push them apart, and the first few eigenvectors of $L_r$ give each node a position in a low-dimensional space. The paper reads the first principal component of that position as a user's attitude toward a topic. Correspondence Analysis (CA), a singular-value decomposition of the standardized residual matrix of rows (users) against columns (outlets or other users), supplies the positive-only counterfactual. To prove that the signed-unsigned gap comes from the negative edges rather than the algorithm choice, the paper constructs a SHEEP null model that replaces every missing edge in the unsigned network with an artificial negative weight, mimicking CA's implicit assumption that non-interaction means dissimilarity; the null embedding tracks CA almost perfectly, confirming that the real negative edges are the carrier of the new information.","core_discovery":"The paper's central discovery, stated on its own terms, is that the signed SHEEP embedding and the unsigned Correspondence Analysis are near-identical in their bulk ranking of users and news outlets—Spearman correlations of 88% on the Russia-Ukraine topic and 80% on general politics—yet they disagree precisely where it matters: the tails of the attitude distribution. In the Russia-Ukraine network, SHEEP places users who upvote Russia Today, Diario Octubre, and Actualidad RT at the hard extreme of the anti-NATO faction, while CA maps those same users to the same ideological level as left-wing outlets like ctxt and Público. In the Broad Politics network, only SHEEP separates far-left users who cast negative votes against the far right from other left-wing users. The paper concludes that negative ties are the signal that reveals which users are not merely opposed in opinion but actively antagonistic toward the opposing faction, and it verifies that the difference comes from the edges themselves: a null model that feeds CA's implicit coding of absent edges as dissimilarity into SHEEP makes the two embeddings converge (Pearson correlations of 0.83 for Russia and 0.94 for politics).","pith_inferences":["Because the signed and unsigned networks differ mainly in how missing edges are coded, the 'hiddenness' of the extremes is conditional on Menéame's interaction semantics: on a platform where absence signals disagreement, positive-only data could recover the same tails, and the paper's achievement would be a coding correction rather than a new social fact.","The finding that far-left users downvote the far right while far-right users stay isolated is presented as platform-specific; a testable generalization for other signed platforms (for instance, Reddit communities with visible downvotes) is that the direction of cross-faction negative voting follows the platform's ideological skew rather than any universal asymmetry.","The signed embedding's tail membership gives unlabeled data a usable signal: the tails could serve as pseudo-labels for training classifiers on positive-only features, testing whether antagonism can be predicted from interaction volume and structural position or whether the vote sign itself is irreducible."],"forward_implications":["On platforms that record only positive engagement (likes, retweets, upvotes), structural-polarization maps will systematically mis-rank the most hostile users: the extremes of the embedding are occupied by negative-vote-heavy users only when downvotes are part of the input.","For the Ukraine-war discussion specifically, the signed network separates the RT-upvoting clique as an extreme faction, so any monitoring or moderation built on positive interactions alone would fail to flag the users most associated with pro-Kremlin content on Menéame.","SHEEP and CA agree strongly on bulk ideology, but the SHEEP null model (missing edges recoded as negative) reproduces CA nearly perfectly (Pearson correlations 0.83 and 0.94), pinning the signed-unsigned gap on the real negative edges rather than on the choice of embedding algorithm.","Only SHEEP separates far-left users who cast negative votes against the far right, and the k-means faction analysis shows these cross-faction negative votes are concentrated in the extremal bins, so signed embeddings change not just who is ranked where but the measured direction of antagonism between factions."],"supporting_citations":[{"why":"Supplies the SHEEP spectral embedding over the repelling Laplacian that produces the signed-network positions of users and outlets.","marker":"[51]"},{"why":"Supplies Correspondence Analysis, the positive-only embedding used as the counterfactual throughout the study.","marker":"[59]"},{"why":"Provides the latent-space assumption linking interaction to ideological distance and the Twitter-based methodology for validating outlet positions.","marker":"[16]"},{"why":"Grounds the interpretation of observed negative ties as signals of discord and tension between users.","marker":"[33]"},{"why":"Earlier signed-network study of a polarized debate used as the template claim that negative ties reveal hidden antagonisms.","marker":"[40]"},{"why":"Ideal-points spatial voting model that justifies treating positive and negative votes as a function of latent ideological difference.","marker":"[26]"},{"why":"BERTopic topic model used to label the Russia-Ukraine and Broad Politics corpora from story descriptions.","marker":"[62]"},{"why":"hSBM topic model used to cross-validate the BERTopic macro-topics before building the topic-specific networks.","marker":"[63]"}],"fun_headline_variants":["Downvotes expose extremists invisible to positive-only maps","Signed ties unmask hidden extremes in polarized networks","Negative edges reveal the most extreme users on Menéame","Only negative ties can tag the hardest-core partisans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that negative ties reveal hidden extremes rests on coding an absent interaction as a neutral zero-weight edge in the signed network; if non-interaction between two users actually signals disagreement, then positive-only methods would already recover the extremes and the contrast the paper draws would be an artifact of that coding choice.","fun_headline_variants_meta":{"raw":{"variants":["Downvotes expose extremists invisible to positive-only maps","Signed ties unmask hidden extremes in polarized networks","Negative edges reveal the most extreme users on Menéame","Only negative ties can tag the hardest-core partisans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000322,"raw_usage":{"total_tokens":1815,"prompt_tokens":952,"completion_tokens":863,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":800}},"tokens_in":568,"tokens_out":863,"duration_ms":6376,"temperature":1.0,"reasoning_tokens":800,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:13:11.334882+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Russia-Ukraine user network with one change: give the zero-weight missing edges a small negative weight calibrated to the observed downvote base rate on Menéame rather than treating them as neutral, and re-compute the SHEEP extremes. If the RT-upvoting clique then stops being uniquely extreme, or if CA recovers the same tails under that coding, the paper's central distinction collapses. A complementary check: among user pairs with no recorded interaction, look at the sign of the first vote they later exchange; a predominantly negative sign would mean absence itself carries a whiff of dissent, contradicting the neutrality assumption the result depends on.","supporting_citations":[{"cited_title":"Sheep, a signed hamiltonian eigenvector embedding for proximity.Communications Physics, 7(1):8, 2024","cited_arxiv_id":null,"evidence_quote":"Supplies the SHEEP spectral embedding over the repelling Laplacian that produces the signed-network positions of users and outlets."},{"cited_title":"chapman and hall/crc, 2017","cited_arxiv_id":null,"evidence_quote":"Supplies Correspondence Analysis, the positive-only embedding used as the counterfactual throughout the study."},{"cited_title":"Jost, Jonathan Nagler, Joshua A","cited_arxiv_id":null,"evidence_quote":"Provides the latent-space assumption linking interaction to ideological distance and the Twitter-based methodology for validating outlet positions."},{"cited_title":"Social Science–guided Feature Engineering: A Novel Approach to Signed Link Analysis.ACM Transactions on Intelligent Systems and Technology, 11(1):1– 27, February 2020","cited_arxiv_id":null,"evidence_quote":"Grounds the interpretation of observed negative ties as signals of discord and tension between users."},{"cited_title":"Why it is important to consider negative ties when studying polarized debates: A signed network analysis of a dutch cultural controversy on twitter.PloS one, 16(8):e0256696, 2021","cited_arxiv_id":null,"evidence_quote":"Earlier signed-network study of a polarized debate used as the template claim that negative ties reveal hidden antagonisms."},{"cited_title":"Poole and Howard Rosenthal","cited_arxiv_id":null,"evidence_quote":"Ideal-points spatial voting model that justifies treating positive and negative votes as a function of latent ideological difference."},{"cited_title":"Peixoto, and Eduardo G","cited_arxiv_id":null,"evidence_quote":"hSBM topic model used to cross-validate the BERTopic macro-topics before building the topic-specific networks."}],"review_version":1}