{"id":"6e55dfa9-989b-4e8e-bd1e-a8ea857806af","arxiv_id":"2411.15458","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"TANGNN combines local neighbor aggregation with a global top-m attention mechanism and introduces ArXivNet, a citation graph with positive, neutral, and negative sentiment labels.","lead":"This paper proposes a graph neural network that combines local neighbor aggregation with a top-m attention mechanism and introduces ArXivNet, a citation network with positive, neutral, and negative sentiment labels. The authors report gains over existing models on classification, link prediction, sentiment prediction, and graph regression, but the evaluation lacks error bars and the attention routine may not do what the text claims.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Top-m attention described in §4.2.2–4.3 is not a per-node similarity-based selection: ranking by a single global auxiliary vector a either yields the same set for all nodes or a rank-neighborhood, and Eqs. (1)–(4) never define a query vector for the central node v.","rationale":"The reader's verdict of REJECT rests primarily on the gap between the claimed per-node Top-m attention and the anchor-vector implementation. My stress-test confirms this is the most load-bearing issue: the abstract and §4.1 promise node-specific selection of the most relevant distant nodes, while §4.3 describes a global ranking by a single vector a, and the attention equations are ambiguous about who the query belongs to. This is not a cosmetic wording problem; it affects what the model actually computes and why it would help. The empirical evaluation has additional problems (no error bars, nonstandard splits, unvalidated ArXivNet labels), but those would only change the strength of the evidence, not the identity of the proposed method. The code is publicly available, so the central ambiguity is directly checkable. If the implementation turns out to use a per-node rank-neighborhood or pair-wise similarity, the paper still needs to validate the transitivity assumption and fix the query-definition issue; under the current text, the mechanism is mis-specified. Therefore the reader's REJECT verdict remains appropriate, pending the code inspection.","tokens_in":21722,"tokens_out":9463,"duration_ms":80166,"concrete_test":"Inspect the released code at github.com/ejwww/TANGNN to locate the Top-m function called by Algorithm 3, and run a diagnostic on Cora after training: for 1000 random nodes, compare the model's selected top-m sets against the true per-node top-m sets obtained by pair-wise cosine similarity on the penultimate-layer representations. If all nodes share the same set (global top-m) or if the mean Jaccard overlap between the model's sets and the true per-node sets is below 0.5, the efficient algorithm does not implement the claimed per-node Top-m attention; additionally, check whether the query in Eq. (1) is computed for the central node or only for nodes inside H_i_Top-m.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (abstract, §1, §4.1) is that each node v aggregates information from the Top-m nodes most similar to v. The efficient algorithm in §4.3 instead computes s_n = a^T ĝ_n for every node with one global vector a, sorts all nodes into one list S, and then derives the Top-m from that list. If the first m entries of S are used, every node receives the same H_i_Top-m, so the attention output is identical for all nodes and the per-node claim collapses. If instead the intended reading is that each node takes its rank-neighbors in S, then the selected nodes are those with similar projections onto a, not necessarily the most similar in feature space; the 'transitive property of similarity' is an unvalidated heuristic that can fail badly for high-dimensional representations. Independently, Eqs. (1)–(4) compute query, key, and value all from H_i_Top-m; there is no query for the central node v unless v is itself in the selected set, which the algorithm does not guarantee. The described per-node attention to relevant distant nodes is therefore either undefined or not what is implemented, so the claimed novelty and advantage over graph transformers are unsupported by the paper's own specification.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TANGNN, a GNN layer that combines a sampled neighborhood aggregation component with a 'Top-m attention' component. The stated goal is to give each node a receptive field consisting of both its local neighbors and the m nodes most similar to it, without the quadratic cost of full graph transformers. An efficient variant replaces all-pairs similarity with a single auxiliary vector a and a sorting step, and a layer-concatenation variant TANGNN-LC is also presented. The authors introduce ArXivNet, a citation-sentiment dataset, and report experiments on node classification, link prediction, sentiment prediction, graph regression, and visualization against GCN, GraphSAGE, GAT, GIN, JK-Net, and several graph transformer baselines. The main claims are that TANGNN expands the receptive field while remaining scalable, and that it outperforms existing methods across multiple tasks.","tokens_in":21993,"tokens_out":8736,"duration_ms":74259,"significance":"If the mechanism worked as described, TANGNN would be an interesting intermediate point between local GNNs and global graph transformers, and ArXivNet would be a useful resource for citation-sentiment analysis. The paper also releases code and data, which is a concrete asset, and I see no circular reasoning: the choice of m is a hyperparameter selected by sensitivity analysis. However, the implementation in Section 4.3 does not realize the per-node similarity-based selection promised in the abstract and Section 4.1, and the reported experiments are too weak to establish the claimed superiority. The significance of the contribution therefore rests on correcting the mechanism and substantially strengthening the evaluation.","major_comments":[{"comment":"The efficient Top-m algorithm is not the per-node attention mechanism described in Section 4.1. Eq. (10) computes a single score s_n = a^T ĝ_n for each node against one global auxiliary vector a, and Eq. (11) produces one sorted list S. If the first m entries of S are used, every node receives the same H_i_Top-m; if instead each node takes its rank-neighbors in S, the selected nodes are those with similar projections onto a, not the nodes most similar to v in feature space. In neither reading is the per-node Top-m selection of Section 4.1 and Figure 1 implemented. Moreover, Eqs. (1)-(4) derive query, key, and value matrices all from H_i_Top-m, so there is no query vector for the central node v; the attention weight A_uv in Eq. (4) is undefined as a node-specific weight unless v is itself in the selected set, which the algorithm does not guarantee. The 'transitive property of similarity' invoked to justify the anchor-vector proxy is an unvalidated assumption that can fail in high-dimensional spaces.","section":"§4.3, §4.2.2, Algorithm 3"},{"comment":"The experimental evidence is reported as single runs without standard deviations, confidence intervals, or significance tests, and Section 5.1.2 fixes the learning rate, batch size, and depth uniformly for all baselines rather than tuning each method. The differences in the tables are often small, for example Table 3 at 10% training gives TANGNN-LC 0.9657 versus SGFormer 0.9570, which could easily be within run-to-run noise. As presented, the data do not support the abstract's claim that the method 'outperforms existing methods.'","section":"§5.1.2, Tables 3-8"},{"comment":"The evaluation protocol is inconsistent across tasks and datasets: node classification uses AUROC on Cora and Citeseer under arbitrary 10% to 90% training splits (Tables 3 and 4), F1-micro on PubMed (Table 5), and accuracy on ArXivNet (Table 6), while graph regression uses MAE (Tables 7 and 8). For standard transductive benchmarks like Cora and Citeseer, AUROC with varied training fractions is not the usual protocol and makes comparison with prior work difficult. A consistent protocol with standard splits and the appropriate per-task metric is needed for the claimed cross-task superiority.","section":"§5.1.1, §5.2.1, Tables 3-6"},{"comment":"The loss function in Eq. (7), L(p,q) = -sum(p log q + (1-p) log(1-q)), is inconsistent with Algorithm 4, where p ← softmax(g_v) and q ← one_hot(y). If q is a one-hot vector, log q is undefined for zero entries, and the text states that p is the desired output while q is the actual output, reversing the assignments in Algorithm 4. This makes the training objective ambiguous and prevents reproduction of the training procedure as specified.","section":"§4.2, Algorithm 4, Eq. (7)"}],"minor_comments":[{"comment":"The baseline list mentions 'SAGEFormer,' but no such model appears in the results tables; this is likely a typo for 'SGFormer' and should be corrected.","section":"§5.1.2"},{"comment":"Figure 5 varies a parameter K without defining it, while Section 5.1 fixes the number of layers L to 2; the relationship between K and L should be clarified.","section":"§5.2.1, Figure 5"},{"comment":"The reported percentage improvements are not consistent with the tables: on Cora at 50% training, TANGNN-LC gives 0.9735 versus GraphSAGE 0.9056, an absolute gain of about 6.8 points rather than 12%, and on PubMed at 10% training the gain over TANGNN is about 0.4 points rather than 2%.","section":"§5.2.1, Tables 3 and 5"},{"comment":"There are typographical errors in the orthogonality proof, including '∥¯g2∥' instead of '∥¯g∥^2', and the statement that feature vectors with high similarity are mapped close to a after orthogonalizing to the mean is not generally true and should be justified or removed.","section":"§4.3, Eq. (9)"},{"comment":"There are numerous formatting and spelling errors, including 'Dateset' in Table 1, 'ZIN C' in Section 5.1.1, 'GAT 07105' in Table 4, '01685' in Table 8, and 'Sigmod' instead of 'Sigmoid' in Section 4.2.1; a careful proofread is needed.","section":"Tables 1-8"}],"recommendation":"reject","confidential_remarks":"The paper has a useful new dataset and a suggestive idea, but the central mechanism as specified is not the per-node attention mechanism that motivates the work, and the experimental evaluation is not strong enough to support the empirical claims. The issues are load-bearing rather than cosmetic: Section 4.3's global anchor-vector ranking either gives every node the same Top-m set or selects nodes by a stated transitivity heuristic that is not validated, and the attention equations do not define a query for the central node. The reported single-run results and untuned baselines would not convince a careful reader even if the architecture were corrected. I therefore recommend rejection, though I do not see evidence of deliberate circularity in the hyperparameter choice."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the central mechanism is not implemented as advertised. Section 4.3 selects the top-m nodes by cosine similarity to a single global auxiliary vector a, so every node either shares the same top-m set or takes rank-neighbors in a global ordering. Neither matches the per-node \"most similar to v\" story in Section 4.1 and the abstract. And Eqs. (1)-(4) compute queries, keys, and values all from the same H_i_Top-m set, so there is no query for the central node v unless v is itself in the selected set. The stress-test note is right; I verified it against the text.\n\nWhat is genuinely useful: the anchor-vector approximation to all-pairs similarity is a reasonable efficiency trick, and combining it with GraphSAGE-style sampling is a sensible architecture direction. The paper also ships code and a new dataset (ArXivNet) for citation sentiment prediction, which is a real, if small, contribution.\n\nThe soft spots beyond the mechanism mismatch: all tables report single runs with no error bars; baselines are given fixed hyperparameters rather than tuned; Cora and Citeseer are evaluated with AUROC under arbitrary 10-90% splits; and the ArXivNet sentiment labels are produced by another model (DictSentiBERT) with no manual validation or agreement numbers. The parameter m is chosen by sensitivity analysis, which is fine, but the lack of repeated trials means the claimed margins over SGFormer and NAGphormer (often a few points) are not credible as stated.\n\nWho this is for: someone working on efficient GNN attention or citation sentiment could read this for the dataset and the anchor-vector trick, but they would need to re-implement the mechanism from scratch because the paper's own specification is ambiguous. It deserves a serious referee—it is not a desk reject—but the referee should ask for a corrected algorithm description and a much more careful evaluation. My own verdict would be reject in current form, with a path to conditional acceptance if the mechanism is fixed and the experiments are redone with multiple seeds and tuned baselines.","headline":"The top-m mechanism as implemented is not the per-node attention the paper claims, the experiments are too weak to rescue it, but the ArXivNet dataset and the anchor-vector efficiency trick are salvageable.","tokens_in":22526,"tokens_out":3480,"would_cite":false,"duration_ms":31746,"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":"TANGNN adds a Top-m attention branch to neighborhood aggregation, giving each GNN layer both local and long-range information at near-linear cost.","keywords":["graph neural networks","top-m attention","neighborhood aggregation","graph representation learning","link prediction","citation sentiment prediction","graph regression","scalable graph learning"],"falsifier":"Run TANGNN on a graph where the global top-m set differs sharply from the per-node top-m set—for example, two dense clusters whose feature directions are orthogonal, so each cluster's most relevant nodes are inside the cluster but the global ranking is dominated by one cluster—and compare TANGNN's accuracy with a variant that computes the true per-node top-m. If the global-ranking model loses, the auxiliary-vector shortcut, not per-node attention, is what the implementation actually uses; inspecting the Top-m function in the released code to see whether it returns a per-node subset or the same sorted list S for all nodes would settle the question directly.","tokens_in":21485,"feed_emoji":"🔗","tokens_out":7494,"duration_ms":63429,"temperature":0.7,"pith_summary":"TANGNN is a graph neural network architecture designed to give each node a wide receptive field without the quadratic cost of graph transformers. It runs two aggregation branches in every layer: a standard sampled-neighborhood aggregation and a Top-m attention branch that selects the m most similar nodes and applies self-attention to them. The paper claims this two-branch design outperforms established GNNs and graph transformers on vertex classification, link prediction, citation-sentiment prediction, graph regression, and visualization, while running faster on large graphs. For the sentiment task it introduces ArXivNet, a citation network labeled positive, neutral, or negative, which the authors say is a first for GNN research. If the empirical claims hold, TANGNN offers a practical middle path between shallow local GNNs and expensive global attention models.","feed_headline":"TANGNN adds Top-m attention to GNNs to reach distant nodes cheaply","feed_subtitle":"Two branches, sampled neighbors and top-m attention, give faster convergence and higher accuracy than graph transformers, the paper reports.","key_machinery":"The load-bearing object is the Top-m efficient algorithm: an auxiliary vector a, kept orthogonal to the mean node vector by the update $a = \\mathrm{L2Norm}(a - (a^T \\bar{g})\\bar{g})$, scores every node by cosine similarity $s_n = a^T \\hat{g}_n$, and sorting these scores produces one global list S whose top m entries form the attention set. This turns all-pairs similarity from $O(N^2)$ to $O(N)$. The other mechanism is the neighborhood aggregation branch, which samples a fixed number of neighbors and applies a mean aggregator in a GraphSAGE-style pattern. Each layer concatenates the two branch outputs and sends them through an MLP; TANGNN-LC additionally concatenates outputs across layers so the final representation keeps both shallow and deep information.","core_discovery":"On the paper's own terms, the central discovery is that a GNN layer can combine local message passing with a cheap global-attention-style component by selecting the Top-m most similar nodes and applying scaled dot-product attention only within that small set. Similarity is not computed between all node pairs; instead each node's embedding is scored against one learnable auxiliary vector a using cosine similarity, the scores are sorted once, and the top m entries become the attention set. The two branches, neighborhood aggregation and Top-m attention, are concatenated and passed through an MLP at each layer, and the LC variant concatenates all layer outputs before a final MLP. The paper reports that the resulting model and its variants beat GCN, GraphSAGE, GAT, GIN, JK-Net, Graphormer, TransGNN, NAGphormer, SAT, DeepGraph, and SGFormer on the tasks it evaluates, and that the Top-m sampling keeps memory and runtime low enough for datasets where some graph transformers run out of memory.","pith_inferences":["The paper describes the Top-m set as node-specific, but its Section 4.3 implementation sorts all nodes by $a^T \\hat{g}_n$ once and reuses the same sorted list for every node; the authors leave implicit that the attention branch is therefore aggregating a global set of prototype-similar nodes rather than per-node nearest neighbors.","A testable consequence is that TANGNN's gains could come from a shared global context rather than pairwise relevance, which would connect it to global-memory or graph-pooling designs; comparing it with a true per-node top-m variant would separate those explanations.","The auxiliary-vector transitivity assumption is the natural extension point: on graphs with strong cluster structure, a single direction a cannot represent all relevance relations, so a multi-vector or per-cluster extension is a straightforward experiment."],"forward_implications":["Graph models can expand their receptive field to distant nodes without stacking many layers, which is the setting where oversmoothing typically degrades GNN accuracy.","Because similarity scoring is $O(N)$ and both branches sample a fixed number of nodes per layer, the model's memory and runtime scale better than graph transformers on large graphs; the paper shows Graphormer and SAT overflowing memory on Reddit while TANGNN converges.","The TANGNN-LC variant, which concatenates layer outputs, gives a direct way to keep early-layer local information in the final embedding, which the paper links to its best classification results.","Citation sentiment prediction on ArXivNet becomes a new benchmark task for GNNs, with edges labeled positive, neutral, or negative rather than only nodes or whole graphs."],"supporting_citations":[{"why":"supplies the sampled-neighborhood aggregation pattern used by the neighbor branch.","marker":"Hamilton et al., 2017"},{"why":"supplies the scaled dot-product self-attention used inside the Top-m branch.","marker":"Vaswani et al., 2017"},{"why":"supplies the auxiliary-vector similarity trick that reduces all-pairs scoring to O(N).","marker":"Zhuo et al., 2021"},{"why":"Graphormer is a primary transformer baseline with structural encodings that TANGNN must beat.","marker":"Ying et al., 2021"},{"why":"TransGNN is a transformer-GNN hybrid baseline that the paper compares against.","marker":"Zhang et al., 2023"},{"why":"SGFormer is an efficient single-layer-attention transformer baseline in the comparison.","marker":"Wu et al., 2024"},{"why":"SPECTER provides the semantic node features used to build ArXivNet.","marker":"Cohan et al., 2020"},{"why":"DictSentiBERT supplies the sentiment labels for the citation-sentiment task.","marker":"Yu & Hua, 2023"},{"why":"unarXive is the source of papers used to construct the ArXivNet citation network.","marker":"Saier et al., 2023"}],"fun_headline_variants":["Top-m attention makes GNNs reach distant nodes cheaply","TANGNN: scalable graph learning with top-m attention","New GNN beats graph transformers via top-m attention","Top-m attention GNN boosts efficiency and accuracy","Scalable GNN with top-m attention outperforms on graphs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that ranking every node by cosine similarity to one shared auxiliary vector a correctly identifies the nodes most relevant to each individual node, even though the ranking is the same for all nodes.","fun_headline_variants_meta":{"raw":{"variants":["Top-m attention makes GNNs reach distant nodes cheaply","TANGNN: scalable graph learning with top-m attention","New GNN beats graph transformers via top-m attention","Top-m attention GNN boosts efficiency and accuracy","Scalable GNN with top-m attention outperforms on graphs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000221,"raw_usage":{"total_tokens":1492,"prompt_tokens":1029,"completion_tokens":463,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":383}},"tokens_in":645,"tokens_out":463,"duration_ms":4789,"temperature":1.0,"reasoning_tokens":383,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:16:46.982622+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run TANGNN on a graph where the global top-m set differs sharply from the per-node top-m set—for example, two dense clusters whose feature directions are orthogonal, so each cluster's most relevant nodes are inside the cluster but the global ranking is dominated by one cluster—and compare TANGNN's accuracy with a variant that computes the true per-node top-m. If the global-ranking model loses, the auxiliary-vector shortcut, not per-node attention, is what the implementation actually uses; inspecting the Top-m function in the released code to see whether it returns a per-node subset or the same sorted list S for all nodes would settle the question directly.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"supplies the sampled-neighborhood aggregation pattern used by the neighbor branch."},{"cited_title":"Graph Neural Networks with Feature and Structure Aware Random Walk","cited_arxiv_id":"2111.10102","evidence_quote":"supplies the auxiliary-vector similarity trick that reduces all-pairs scoring to O(N)."},{"cited_title":", author Cai, T","cited_arxiv_id":null,"evidence_quote":"Graphormer is a primary transformer baseline with structural encodings that TANGNN must beat."},{"cited_title":"TransGNN: Harnessing the Collaborative Power of Transformers and Graph Neural Networks for Recommender Systems","cited_arxiv_id":"2308.14355","evidence_quote":"TransGNN is a transformer-GNN hybrid baseline that the paper compares against."},{"cited_title":", author Krause, J","cited_arxiv_id":null,"evidence_quote":"unarXive is the source of papers used to construct the ArXivNet citation network."}],"review_version":1}