{"id":"63958655-cdaa-4448-9283-1821aa710a30","arxiv_id":"2502.00351","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MOHGCAA, a multi-order hyperbolic graph convolution with aggregated attention, is reported to outperform prior Euclidean and hyperbolic baselines on four datasets in supervised and unsupervised settings.","lead":"This paper introduces MOHGCAA, a graph neural network that combines multi-order neighborhood aggregation with hyperbolic embeddings for social event detection. The authors report consistent but small gains over prior hyperbolic and Euclidean baselines, though no code or error bars are provided.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 9's attention weight v_i^k is independent of order k, so the claimed adaptive multi-order aggregation is not implemented as written; this undercuts the central novelty and the attribution of the reported gains.","rationale":"The reader's weakest assumption identifies exactly the most load-bearing point: Eq. 9 does not implement order-dependent attention. I checked the equation and the surrounding text; the k-independence is concrete, not a matter of interpretation. This is not an accusation of dishonesty; the likely explanation is a missing subscript or an incorrect softmax argument, but as published the central mechanism is absent. The paper's empirical claim could still survive if the code uses a different attention, which is why the appropriate disposition remains conditional rather than outright rejection. The decisive check is implementation inspection plus the proposed ablation. Secondary issues such as the reversed exp/log descriptions in Section 4.3.1 and the undefined kappa_ij and adjacency powers should also be fixed, but Eq. 9 is the single load-bearing defect.","tokens_in":17039,"tokens_out":6864,"duration_ms":69971,"concrete_test":"Request or reconstruct the implementation and inspect the forward pass that computes Eq. 8-9: determine whether v_i^k is a function of h_i^k or of the order index k. Then run an ablation that replaces the implemented attention with the paper's Eq. 9 (order-independent weights). If performance is unchanged, the claimed adaptive aggregation is not the source of the reported gains; if performance drops, the paper's equation is incorrect and must be amended and re-evaluated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1.4 defines the aggregated attention as h_i = sum_{k=1}^K v_i^k h_i^k (Eq. 8), with v_i^k = softmax(s_i^k) = exp(W_V tanh(W x_i)) / sum_{j=1}^N exp(W_V tanh(W x_j)) (Eq. 9). As written, s_i^k does not depend on order k, on h_i^k, or on the order-specific representation; it is the same scalar for every k of a given node i. The denominator runs over nodes, not over the K orders, so the weights are not normalized across orders. Consequently the mechanism cannot highlight the most relevant orders for each node's representation as claimed in Section 4.1.4. If Eq. 9 reflects the implementation, the model reduces to a fixed or arbitrarily scaled combination of multi-order features, removing the adaptive aggregation that is the paper's central architectural contribution and the stated reason for the reported SOTA numbers. If the implementation differs, the paper's formal description is wrong and cannot be used to reproduce or evaluate the method. No code or data are provided to resolve this. The same section also leaves kappa_ij (Eq. 7) and the multi-order adjacency powers a^k undefined, but Eq. 9 alone is sufficient to block the central claim as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes MOHGCAA, a graph-convolution framework for social event detection that combines multi-order adjacency powers with hyperbolic-space embeddings and an attention-based aggregation across orders. The authors present two instantiations: an unsupervised contrastive version evaluated on mini-Twitter, Cora, and Citeseer, and a supervised version evaluated on Twitter. They report that MOHGCAA outperforms Euclidean and hyperbolic baselines on Micro-F1/Macro-F1 in the unsupervised setting and on ACC/NMI/AMI/ARI in the supervised setting. The claimed contribution is an adaptive attention mechanism that emphasizes the most task-relevant multi-order relational neighborhoods for each node, implemented in tangent space with mappings to and from hyperbolic space.","tokens_in":17297,"tokens_out":7742,"duration_ms":74863,"significance":"If the formal description were corrected, the paper would offer a plausible combination of existing ideas: hyperbolic tangent-space GCNs, multi-order adjacency powers (as in MOGANED), and contrastive pretraining. The empirical evaluation spans three benchmark datasets plus a large Twitter dataset in both supervised and unsupervised regimes, and the claim of consistent state-of-the-art performance is falsifiable. The paper does not provide code, repeated-run statistics, or machine-checked derivations, so the evidence currently rests on point estimates. The central architectural claim--adaptive order-selective attention--is not implemented by the equations as written, which is the main obstacle to accepting the paper's interpretation of its own results.","major_comments":[{"comment":"The attention weight v_i^k does not depend on the order k or on the order-specific representation h_i^k; the score s_i^k = W_V tanh(W x_i) is identical for every k, and the denominator is a sum over nodes rather than over orders. Consequently Eq. (8) reduces to a fixed linear combination of the multi-order features and cannot 'highlight the most relevant orders' as stated in Section 4.1.4 and claimed as the reason for the gains in Section 5.2.4. If the implementation differs from this equation, the formal description must be corrected and the actual aggregation rule specified; if Eq. (9) matches the implementation, the central novelty is not present.","section":"Section 4.1.4, Eq. (9)"},{"comment":"The notation a^k_along, a^k_rev, and a^k_loop is never defined; no formula is given for the k-th power of the adjacency matrices or for their normalization. The curvature score kappa_ij in Eq. (7) is also never defined despite the text saying it is based on Ollivier Ricci curvature. Since alpha_ij enters every convolution order and kappa_ij is its only input, this makes the multi-order convolution non-reproducible as written.","section":"Section 4.1.3, Eqs. (5)-(7)"},{"comment":"The exposition reverses the exponential and logarithmic maps. The text says Euclidean features are mapped into hyperbolic space using the logarithmic map and then projected onto the tangent space via the exponential map, and that final features are mapped back into hyperbolic space using the logarithmic map. This contradicts Eqs. (23)-(24) and Section 4.1.5, where the correct order is exp (Euclidean to hyperbolic), log (hyperbolic to tangent), and exp (tangent back to hyperbolic). The section should be rewritten for consistency.","section":"Section 4.3.1"},{"comment":"The results are reported as point estimates only, with no standard deviations, number of runs, or significance tests. Several reported improvements are small in absolute terms (e.g., Cora Micro-F1 0.8543 vs. 0.8314 for UHSED; Twitter ACC 0.91 vs. 0.89 for HNN and HSED), so the claim in Sections 5.2.4 and 5.3.4 that MOHGCAA 'consistently outperforms' the baselines is not statistically supported. The authors should report repeated-run statistics or perform significance testing.","section":"Tables 4 and 6"},{"comment":"The number of multi-order steps K (4 for unsupervised, 2 for supervised) is selected after inspecting performance on the evaluation data, and the final Tables 4 and 6 use the selected configuration. If K is chosen on the same test data, the reported comparisons are optimistically biased. The authors should specify a validation-based selection procedure or report results for all K on a hold-out set.","section":"Sections 5.2.5 and 5.3.5"}],"minor_comments":[{"comment":"The heading 'Relate Work' should be 'Related Work'.","section":"Section 2 heading"},{"comment":"The row name 'MOHGAA' is inconsistent with the model name 'MOHGCAA' used elsewhere in the paper.","section":"Table 4"},{"comment":"The 'Kelin model' (also written 'Kälin Model' in Section 3.2) should be the Klein model.","section":"Table 1"},{"comment":"'Traing rate' should be 'Training rate'.","section":"Table 5"},{"comment":"DGI is attributed to reference [43], which is the Graph Attention Networks paper; DGI should instead cite Veličković et al., Deep Graph Infomax.","section":"Section 5.2.1"},{"comment":"The caption states that the final representation is mapped into hyperbolic space through the log_o function; based on the pipeline it should be exp_o.","section":"Figure 5 caption"},{"comment":"The text says the node set changes (V ≠ V') while the adjacency matrix remains the same; feature corruption typically keeps the node set identical and changes feature values. Please clarify.","section":"Section 4.2.2"},{"comment":"The contrastive loss formula appears garbled in the manuscript text, making it impossible to verify the objective; please re-typeset and define all symbols.","section":"Section 4.2.4, Eq. (20)"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a social-computing or applied graph-learning venue, and the empirical breadth is a strength. However, the formal inconsistencies in the attention mechanism and the undefined curvature/adjacency notation are load-bearing for the central claim, and the lack of variance estimates weakens the empirical conclusion. If the authors can correct the equations, define all notation, add repeated-run statistics, and clarify the model-selection protocol, the paper could become publishable. I would also recommend asking the authors to state explicitly how the proposed method differs from MOGANED [28] beyond the use of hyperbolic space."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible combination of known components—MOGANED's multi-order graph convolution and aggregated attention, dropped into hyperbolic space via HGCN/HNN-style maps—and the reported numbers beat the listed baselines on every dataset. The unbiased part: the high-level idea is reasonable, and the ablation study on multi-order steps and dimensions is a useful check. But the paper has load-bearing formal problems.\n\nThe biggest one is Eq. 9. The attention weight v_i^k is written as softmax(s_i^k), but s_i^k depends only on x_i, not on k or on the order-specific h_i^k. The denominator runs over all N nodes, not over the K orders, so the weights aren't even normalized across orders. As written, the model cannot adaptively emphasize the most relevant order for a node, which is exactly the contribution claimed in Section 4.1.4. Either the equation is a typo and the implementation does something else, or the claimed mechanism isn't what's running. Either way, the paper's formal description doesn't support its central novelty. That alone blocks the current version.\n\nThere are smaller issues in the same spirit. Section 4.3.1 says the node features are mapped into hyperbolic space with the logarithmic map and then projected to the tangent plane with the exponential map, which is backwards (Section 4.1.2 gets it right). The curvature score kappa_ij in Eq. 7 is listed as 'a learnable score' but never defined. And the multi-order adjacency powers a^k along/rev/loop are never specified—are they graph powers, or something like MOGANED's sub-matrices? That matters for reproduction.\n\nEmpirically, the numbers look good, but there are no error bars or significance tests, and no code. The most direct predecessor, MOGANED, is cited but never benchmarked. Given that the multi-order convolution and attention are inherited from MOGANED, omitting it as a baseline makes the comparison incomplete. The gains over the strongest hyperbolic baselines are about 2-3 points, so the practical significance is real but modest.\n\nWho gets value: people working on hyperbolic GNNs for event detection will find the combination worth considering once the formal bugs are fixed. The paper deserves a serious referee—it's not a desk-reject—but the referee should send it back for major revision, asking for code, corrected equations, error bars, and a MOGANED comparison. I'd phrase it as 'promising but not ready.'","headline":"The paper is a plausible reuse of MOGANED's multi-order convolution plus hyperbolic mappings, but Eq. 9 removes the claimed adaptive order-weighting and, without code or a MOGANED baseline, the empirical claims can't be checked.","tokens_in":17823,"tokens_out":5663,"would_cite":false,"duration_ms":50933,"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 a multi-order hyperbolic graph convolution with aggregated attention outperforms prior Euclidean and hyperbolic baselines on social event detection and node classification benchmarks.","keywords":["social event detection","hyperbolic space","multi-order graph convolution","attention aggregation","contrastive learning","node classification","graph neural network"],"falsifier":"Implement equation (9) exactly as stated and, in a controlled run, replace the attention score with one that uses the order-specific representation $h_i^k$; if both versions achieve the same performance on mini-Twitter, the reported gains do not come from adaptive order selection. A simpler check is to print the weights $v_i^k$ during training — if they are identical across orders for every node, the mechanism is inert.","tokens_in":16797,"feed_emoji":"📈","tokens_out":8321,"duration_ms":79074,"temperature":0.7,"pith_summary":"Social media messages connected by replies and mentions form tree-like hierarchies that are hard to represent in flat Euclidean space. This paper claims that a graph neural network operating in hyperbolic space, where hierarchical structure is easier to embed, can detect real-world events better than prior Euclidean and hyperbolic methods. The proposed framework, MOHGCAA, runs graph convolutions of multiple orders on directed, reverse, and self-loop adjacency matrices in the tangent plane of a hyperbolic model, then combines the order-specific features with an attention-weighted sum before mapping back into hyperbolic space. In supervised and unsupervised experiments on Twitter, mini-Twitter, Cora, and Citeseer, it reports consistently higher scores than the compared baselines. If the results hold, the framework offers a general recipe for hierarchical graph learning beyond social media.","feed_headline":"Hyperbolic graph net tops social event detection benchmarks","feed_subtitle":"Multi-order attention in hyperbolic space beats prior models on F1 for mini-Twitter, Cora, and Citeseer.","key_machinery":"The load-bearing component is the multi-order graph convolution operator, which uses three adjacency matrices per node — edges in the original direction, reverse direction, and self-loops — to produce order-specific feature representations $h_i^k$. These are aggregated via attention weights $v_i^k$ to form a single node representation $h_i$ in the tangent space of a hyperbolic model, and the result is mapped back with the exponential map. The design is intended to capture higher-order syntactic relationships without stacking deep graph convolution layers, thereby avoiding over-smoothing. The hyperbolic host space is what lets the model represent the tree-like structure of social media data with low distortion.","core_discovery":"The central claim is that modeling higher-order relationships between event messages is the missing ingredient in social event detection, and that doing it in hyperbolic space is better than doing it in Euclidean space. The authors report that MOHGCAA achieves state-of-the-art Micro-F1 and Macro-F1 on mini-Twitter, Cora, and Citeseer in the unsupervised setting, and the best ACC, NMI, AMI, and ARI on the Twitter dataset in the supervised setting, compared with all listed baselines. The discovery is an empirical one: a single architecture that combines multi-order graph convolution, aggregated attention, and hyperbolic embeddings can dominate across both paradigms and across social and general graph datasets.","pith_inferences":["If the attention equation really is order-invariant as written, then the reported improvements likely come from the multi-order convolution itself or the hyperbolic embedding, not from the adaptive order selection the paper claims is central; an ablation that removes the attention layer would test this.","The method's success on citation networks hints that any hierarchical or tree-structured dataset — taxonomies, file systems, knowledge graphs — could benefit from the same multi-order hyperbolic recipe, a possibility the paper does not explore.","A stronger test of the hyperbolic claim would compare against a Euclidean baseline with identical multi-order convolution and attention, controlling for parameter count; the paper's Euclidean counterpart reports lower scores, but the setting may not be perfectly matched."],"forward_implications":["If the reported results hold, social event detection can be improved by combining hyperbolic embeddings with multi-order convolutions, rather than relying on deeper GCN layers.","The unsupervised version's gains on Cora and Citeseer suggest the framework generalizes to non-social graph node classification, making it a candidate general-purpose graph encoder.","Because the aggregation happens in the tangent plane, the architecture avoids deep stacking, which points to an efficient way to capture long-range dependencies without over-smoothing.","The same encoder is used for both supervised and unsupervised objectives, so the framework offers a single representation module that can be plugged into different training regimes."],"supporting_citations":[{"why":"It supplies the multi-order graph convolution with aggregated attention design that the paper adapts to hyperbolic space.","marker":"[28]"},{"why":"It provides the tangent-space hyperbolic convolution machinery used for the encoder.","marker":"[45]"},{"why":"It defines the hyperbolic neural network operations that the model and its HNN baseline rely on.","marker":"[46]"},{"why":"It is the hyperbolic social event detection baseline that MOHGCAA extends and compares against.","marker":"[8]"},{"why":"It gives the contrastive learning objective used in the unsupervised variant.","marker":"[40]"},{"why":"It supplies the Twitter dataset used for the supervised evaluation.","marker":"[41]"},{"why":"It supplies the Cora and Citeseer graphs used for the unsupervised evaluation.","marker":"[42]"}],"fun_headline_variants":["Multi-order hyperbolic graph convolution sets new SED records","Hyperbolic attention with multi-order graph beats event detection baselines","Multi-order hyperbolic GCN excels at social event spotting","Multi-order hyperbolic graph + aggregated attention dominates SED"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole argument depends on the attention mechanism in equation (9) actually weighting different relationship orders differently for each message; as written, the weight depends only on the message's own features, not on the order, so it cannot do that.","fun_headline_variants_meta":{"raw":{"variants":["Multi-order hyperbolic graph convolution sets new SED records","Hyperbolic attention with multi-order graph beats event detection baselines","Multi-order hyperbolic GCN excels at social event spotting","Multi-order hyperbolic graph + aggregated attention dominates SED"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0006,"raw_usage":{"total_tokens":2763,"prompt_tokens":867,"completion_tokens":1896,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":1831}},"tokens_in":483,"tokens_out":1896,"duration_ms":13637,"temperature":1.0,"reasoning_tokens":1831,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T19:20:00.090989+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement equation (9) exactly as stated and, in a controlled run, replace the attention score with one that uses the order-specific representation $h_i^k$; if both versions achieve the same performance on mini-Twitter, the reported gains do not come from adaptive order selection. A simpler check is to print the weights $v_i^k$ during training — if they are identical across orders for every node, the mechanism is inert.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It supplies the multi-order graph convolution with aggregated attention design that the paper adapts to hyperbolic space."},{"cited_title":"Chami, Z","cited_arxiv_id":null,"evidence_quote":"It provides the tangent-space hyperbolic convolution machinery used for the encoder."},{"cited_title":"Ganea, G","cited_arxiv_id":null,"evidence_quote":"It defines the hyperbolic neural network operations that the model and its HNN baseline rely on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It is the hyperbolic social event detection baseline that MOHGCAA extends and compares against."},{"cited_title":"Building a large-scale corpus for evaluating event detection on twitter","cited_arxiv_id":null,"evidence_quote":"It supplies the Twitter dataset used for the supervised evaluation."}],"review_version":1}