{"id":"395f9995-7f42-4609-97ec-43bdcf671c22","arxiv_id":"2412.18720","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"ELISE combines signed personalized message passing with low-rank graph refinement to predict link signs in signed bipartite graphs faster and more accurately than previous GNN methods.","lead":"This paper presents ELISE, a graph neural network that learns node embeddings for signed bipartite graphs, where links between two types of nodes carry positive or negative signs. The method avoids adding extra edges that slow down prior techniques, and tests on four real-world datasets show competitive or better link-sign prediction with much lower compute.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains over signed-bipartite baselines are not significance-tested and are within one standard deviation on two of four datasets, so the central 'outperforms competitors' claim is not yet established.","rationale":"The reader's weakest_assumption concerned the low-rank SVD approximation, but the reader also noted the lack of significance testing. I identify the significance/comparison-completeness issue as the single most load-bearing concern because it directly targets the paper's central empirical claim of outperforming signed-bipartite GNN methods. On the two smaller datasets, the reported improvements over the best competitor are within the reported standard deviations; without significance tests, the improvements could be noise. On the two larger datasets, the signed-bipartite baselines either fail with OOM or are absent, so the claim is not tested against the actual competitors there. This is a concrete, checkable weakness that can be settled by rerunning with significance tests and by attempting reduced-memory baseline runs. The SVD assumption is a legitimate risk, but the ablation study provides empirical support on four datasets, and the joint model generally performs best; the significance issue is more immediate and more directly tied to the headline claim. I also note a secondary concern: Theorem 1's O(m+n) claim treats k as constant, but k is defined as min(|U|,|V|)·r, so k scales with the graph size, making the stated complexity proof inconsistent; this affects the efficiency claim but is less central than the missing statistical evidence. The existing CONDITIONAL verdict remains appropriate, with the additional condition that significance tests and, where possible, reduced-memory baseline comparisons be provided.","tokens_in":19535,"tokens_out":8655,"duration_ms":80731,"concrete_test":"Rerun the five-seed experimental protocol and compute paired one-sided Wilcoxon signed-rank tests (or paired t-tests) on test AUC and Macro-F1 for ELISE versus the best signed-bipartite baseline on Review and Bonanza. If p >= 0.05 for either metric, the 'significantly outperforms' claim is unsupported. Additionally, attempt to run SBGNN and SBGCL on ML-1M and Amazon-DM with a reduced embedding size (e.g., 8) or on a 10% edge subsample; if they still run out of memory, the paper should explicitly restrict the state-of-the-art claim to datasets where the baselines fit, and compare ELISE to the strongest unsigned baselines with a clear caveat.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ELISE outperforms existing GNN-based methods for signed bipartite graphs. On Review, ELISE's AUC is 0.7106±0.0424 versus SBGCL's 0.6960±0.0525; on Bonanza, ELISE's AUC is 0.6838±0.0198 versus LightGCL's 0.6662±0.0226 and SBGNN's 0.6534±0.0332. With five random seeds, these differences are smaller than the reported standard deviations, and the paper reports no significance tests, despite the abstract and contributions stating that ELISE 'significantly outperforms' competitors. Furthermore, on ML-1M and Amazon-DM, the two signed-bipartite baselines SBGNN and SBGCL fail with out-of-memory errors (except SBGNN on Amazon-DM), so the head-to-head comparison with signed-bipartite GNN methods is absent on the largest datasets; the 13.66% AUC improvement on ML-1M is against the unsigned LightGCL. Thus, the 'outperforms competitors' claim currently rests on two small datasets with margins that may be sampling noise and on two larger datasets where the actual signed-bipartite baselines could not be run. This is load-bearing because the paper's primary contribution is an empirical state-of-the-art claim; if the margins are not statistically real or the comparison is incomplete, the headline finding is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ELISE, a GNN-based method for representation learning in signed bipartite graphs, addressing link sign prediction. It contributes two encoders: signed personalized message passing, which extends signed random walk with restart to embedding propagation on the signed bipartite graph without adding edges, and refined message passing, which performs the same propagation on a low-rank SVD approximation of the signed biadjacency matrices. The final node representations are the concatenation of the two encoders' outputs, and link signs are predicted with a two-layer MLP on the concatenated embeddings. The paper also provides time and space complexity analyses, claiming O(m+n) complexity under fixed hyperparameters. The experimental section reports link sign prediction results on four real-world datasets, an ablation study, hyperparameter sensitivity analysis, and efficiency comparisons against SBGNN, SBGCL, LightGCN, and LightGCL.","tokens_in":19845,"tokens_out":4723,"duration_ms":40356,"significance":"If the empirical claim is established, ELISE would be a useful and lightweight alternative to existing signed-bipartite GNNs: it avoids adding same-type edges, requires no per-layer weight matrices or auxiliary losses, and provides a simple, theoretically linear-time message-passing framework. The complexity analysis is transparent, and the method is evaluated on publicly available datasets with standard baselines, which supports reproducibility. However, the central 'significantly outperforms competitors' claim is not yet supported by the reported statistics, because on two of the four datasets the margins over the best baselines are within one standard deviation, and on the two largest datasets the main signed-bipartite baselines fail due to memory limitations. The methodological novelty is moderate, but the contributions are clearly presented and the proposed components are sensible.","major_comments":[{"comment":"The headline claim that ELISE 'significantly outperforms' competitors is not supported by the reported statistics: on Review, ELISE AUC is 0.7106±0.0424 versus SBGCL's 0.6960±0.0525, and on Bonanza, ELISE AUC is 0.6838±0.0198 versus LightGCL's 0.6662±0.0226. In both cases the gap is smaller than the reported standard deviation, and no significance tests are provided. Please report paired significance tests (e.g., paired t-test or Wilcoxon signed-rank test over the five seeds) or confidence intervals, and adjust the abstract and introduction wording accordingly.","section":"§5.2, Table 3"},{"comment":"The signed-bipartite competitors SBGNN and SBGCL run out of memory on the two largest datasets (ML-1M and Amazon-DM), except for SBGNN on Amazon-DM. As a result, on ML-1M and Amazon-DM the comparison against the state-of-the-art signed-bipartite GNNs is absent, and the 13.66% AUC improvement on ML-1M is measured against unsigned LightGCL rather than a signed-bipartite baseline. The claimed state of the art over signed-bipartite methods therefore rests on Review and Bonanza, where the margins are statistically fragile. Please attempt memory-efficient runs of these baselines (e.g., CPU execution, edge sampling, or mini-batch training), or, if this is infeasible with the original implementations, state this limitation explicitly and qualify the claim.","section":"§5.2, Table 3"},{"comment":"Equation (2) contains an apparent typo: the update for the negative embedding M_V^{(l)} includes the injection term c·x_U, whereas the surrounding text states that the initial features X_V are injected, and the corresponding update for M_U in Eq. (3) has no injection term. If this is not a typo but a deliberate design, the asymmetry needs a justification. If it is a typo, please correct it and re-check the associated derivations and algorithm listings. Equation (8) also contains unmatched parentheses after RMP_k calls.","section":"§4.2, Eq. (2)"},{"comment":"The ablation does not uniformly support the claim that jointly learning both components is beneficial: on Bonanza, ELISE-R (without refined message passing) achieves test AUC 0.7235±0.0257, while full ELISE achieves 0.6838±0.0198. The paper attributes this to a validation-test distribution gap, but no evidence is provided for that explanation. Please report the selected hyperparameters for each variant on each dataset and consider a nested model-selection protocol so that the component-wise comparison is not confounded by hyperparameter choice.","section":"§5.4, Table 5"}],"minor_comments":[{"comment":"The Amazon-DM Micro-F1 entry for ELISE is reported as 0.9802±0.9802, which appears to be a typo in the standard deviation; please verify and correct.","section":"Table 3"},{"comment":"The Bonanza negative-edge percentage is written as '0...2%'; this should presumably be '2.0%' (738/36,543).","section":"Table 2"},{"comment":"The text refers to the injection-ratio panel as Figure 3(a), but it is actually panel (b); please correct the cross-reference. The caption also labels the panels ambiguously.","section":"§5.5, Figure 3"},{"comment":"The title of Algorithm 2 contains a typo: 'Singed Personalized Message Passing' should be 'Signed Personalized Message Passing'.","section":"Algorithms 2 and 3"},{"comment":"Reference [15] appears to be a duplicate of reference [14] (both cite the SDGNN paper); the source of the Review dataset should be cited explicitly or with a correct reference.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely within scope for a data-mining or applied ML venue. The main obstacle to acceptance is the gap between the claimed empirical superiority and the evidence: statistical significance is not established, and the signed-bipartite comparison is incomplete on the larger datasets. These issues are fixable with additional experiments and more careful wording, so I do not recommend rejection. I would also gently encourage the authors to make the baseline memory failures more visible as a limitation rather than presenting them only as evidence of efficiency."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The strongest thing about this paper is its efficiency story. ELISE avoids adding same-type edges, uses reordered SVD operations so it never forms a dense matrix, and handles datasets like ML-1M and Amazon-DM that make SBGNN and SBGCL run out of memory on a 24GB GPU. The reported speedups (up to 22.5× in training) are concrete and believable given the design. The complexity analysis is honest and the appendix quantifying the added edges (up to 110× the original) explains why the signed baselines fall over.\n\nWhat is actually new: extending signed random walk with restart from score propagation to bipartite node embeddings, and combining it with a low-rank refined message-passing encoder. The low-rank part is the more interesting contribution because it targets the noise issue directly, and the reordering trick to keep it sparse is well done. The ablations show both components contribute, and the paper honestly reports the Bonanza case where ELISE-R beats the full model on test AUC.\n\nSoft spots: the accuracy claims are oversold. The abstract and contributions say ELISE \"significantly outperforms\" competitors, but there are no significance tests. On Review and Bonanza, the AUC gains over the best signed baseline are within one standard deviation. On ML-1M and Amazon-DM, the signed baselines mostly fail, so the head-to-head comparison is incomplete; the 13.66% improvement on ML-1M is against LightGCL, an unsigned method. The efficiency advantage on large graphs is real, but that should be framed as scalability rather than clean SOTA accuracy. There is also a typo in Equation (2): the injection term should be X_V, not x_U, and the negative update seems to be missing the injection term entirely. Minor, but confusing. No code is released, which matters for an efficiency-focused paper.\n\nI would send this to serious peer review. The method is sensible, the efficiency claims are strong, and the accuracy story can be fixed with proper significance testing and more careful framing. The paper will be more useful after those revisions, but it is worth the referees' time now.","headline":"Solid, practical signed-bipartite GNN with real efficiency gains; the accuracy edge over signed baselines is plausible but not statistically established.","tokens_in":20361,"tokens_out":2205,"would_cite":true,"duration_ms":21952,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A lightweight GNN encoder, ELISE, predicts link signs in signed bipartite graphs more accurately and far faster than prior GNN methods.","keywords":["signed bipartite graphs","link sign prediction","graph neural networks","personalized propagation","low-rank approximation","message passing","balance theory","representation learning"],"falsifier":"Construct or find a signed bipartite dataset where the sign of edges is generated from high-rank structure, for example signs determined by per-node random vectors with no low-dimensional latent factor, then compare ELISE with its own variant that uses only the original-graph encoder: if the full model does not beat the variant, the low-rank assumption fails. The paper's own ablation tables provide the template for this test.","tokens_in":19347,"feed_emoji":"🔗","tokens_out":5222,"duration_ms":47487,"temperature":0.7,"pith_summary":"The paper introduces ELISE, a graph-neural-network method for learning node embeddings in signed bipartite graphs—graphs whose edges run between two types of nodes and carry a positive or negative sign. Its central claim is that ELISE predicts the sign of unseen links more accurately than existing GNN-based approaches while training and running substantially faster, and that it can handle large graphs on which those approaches run out of memory. The method achieves this by replacing the naive message passing of prior work with two complementary encoders: signed personalized message passing, which injects each node's own features at every propagation step to avoid over-smoothing while respecting balance theory, and refined message passing on a low-rank approximation of the graph, which suppresses noisy interactions. A sympathetic reader would care because link sign prediction is the core task in applications such as recommendation, e-commerce trust, and peer review, and the paper's efficiency gains are what make the method usable on large real-world signed bipartite graphs.","feed_headline":"Signed-graph encoder wins on link signs and speed","feed_subtitle":"Two message-passing passes—personalized plus low-rank—beat edge-augmenting GNNs and train in linear time.","key_machinery":"The central object is signed personalized message passing, an embedding-level analogue of the signed random walk with restart (SRWR). For each node type the model maintains positive and negative embedding channels; each layer propagates positive embeddings across positive edges and negative embeddings across negative edges (and the cross terms), then injects the node's own input features with ratio $c$ to keep representations distinguishable. The second mechanism is refined message passing: a truncated SVD of the row-wise normalized signed biadjacency matrices, stored once in a hash map and applied through the reordered product $U\\Sigma(V^\\top X)$, so message passing on the reconstructed low-rank graph costs $O(nkd)$ instead of $O(|U||V|d)$ and suppresses noisy edges. Together the two encoders deliver the paper's claimed $O(m+n)$ time and space complexity when rank, dimension, and depth are fixed.","core_discovery":"ELISE's core claim is that a lightweight, edge-free encoder can outperform edge-augmenting GNNs on signed bipartite link sign prediction. Prior GNN methods for these graphs add edges between same-type nodes to encode balance theory, which bloats the graph and slows training; ELISE instead extends personalized propagation—the signed random walk with restart—to node embeddings, carrying positive and negative embedding channels across signed edges and adding a personalized-feature injection at each layer. This respects balance theory without adding edges and prevents embeddings from collapsing with depth. A second encoder runs the same signed propagation on a rank-k truncated SVD reconstruction of the signed biadjacency matrices, with the matrix multiplications reordered as $U(\\Sigma(V^\\top X))$ for efficiency, to denoise the graph and expose global structure. The final embedding concatenates both encoders' outputs, and the whole model trains with a single binary cross-entropy loss, no per-layer weight matrices, and no auxiliary losses. Experiments on four real-world graphs report higher AUC and Macro-F1 than existing signed-bipartite and unsigned-bipartite GNN baselines, with up to 162.8x faster inference on the largest dataset and with competitors failing from out-of-memory on the larger graphs.","pith_inferences":["The low-rank refinement suggests a testable general principle: for sparse signed bipartite graphs, the sign-relevant signal often sits in a low-dimensional subspace, so truncated SVD acts as a denoiser; applying the same two-channel personalized propagation to other signed graph families, such as signed directed bipartite graphs, is a natural next step the paper does not discuss.","The method's speed is what unlocks scale: since signed-bipartite baselines run out of memory on graphs with millions of edges, ELISE-type encoders could enable sign prediction on much larger e-commerce and review platforms, though the paper only demonstrates four datasets.","The injection ratio $c$ and rank ratio $r$ interact: $c=0.15$ and small $r$ performed best on the tested graphs, but a dataset whose sign signal is spread across many singular directions would likely need a larger $r$; this trade-off is a testable hyperparameter hypothesis, not a claim the paper makes."],"forward_implications":["On the four datasets tested, ELISE reports the best AUC and Macro-F1 among all compared GNN methods, with improvements over the best competitor of up to 2.64% in AUC and 4.96% in Macro-F1.","ELISE trains and infers faster than the baselines on every dataset, and it is the only method that completes on the two largest graphs under GPU memory limits.","Because no same-type edges are added and no per-layer weight matrices or auxiliary losses are used, preprocessing and message passing scale linearly in the number of edges and nodes ($O(m+n)$) with fixed hyperparameters.","The ablation study indicates that both encoders contribute: combining signed personalized message passing with refined message passing outperforms either alone on most datasets.","The signed personalized encoder keeps performance stable as the number of layers grows, whereas the edge-augmenting baseline's accuracy drops, evidence that over-smoothing is mitigated."],"supporting_citations":[{"why":"Defines balance in signed bipartite networks via signed butterflies, the balance-theoretic foundation that ELISE's message passing follows.","marker":"[5]"},{"why":"Supplies the signed random walk with restart equations for score propagation that ELISE extends to node embeddings.","marker":"[18]"},{"why":"Provides the layer-wise aggregation and lightweight no-per-layer-weight design that ELISE adopts.","marker":"[11]"},{"why":"Gives the randomized SVD algorithm used for the efficient low-rank refined graph.","marker":"[8]"},{"why":"The signed-bipartite GNN baseline that adds same-type edges; ELISE must beat it and avoids its added edges.","marker":"[12]"},{"why":"The contrastive signed-bipartite baseline; the main competitor for memory and speed on large graphs.","marker":"[40]"},{"why":"Identifies the over-smoothing failure mode that ELISE's personalized feature injection targets.","marker":"[27]"}],"fun_headline_variants":["ELISE: signed bipartite learning without edge bloat","ELISE: personalized propagation for signed graphs, no extra edges","ELISE: lightweight signed graph encoder outruns heavier GNNs","Signed bipartite links: ELISE beats edge-augmenting GNNs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that a truncated low-rank approximation of the signed biadjacency matrix keeps the structure that determines link signs and that whatever is cut off is mostly noise; if a dataset's sign information lives in the discarded tail of the singular spectrum, the refined encoder will hurt rather than help.","fun_headline_variants_meta":{"raw":{"variants":["ELISE: signed bipartite learning without edge bloat","ELISE: personalized propagation for signed graphs, no extra edges","ELISE: lightweight signed graph encoder outruns heavier GNNs","Signed bipartite links: ELISE beats edge-augmenting GNNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000957,"raw_usage":{"total_tokens":4160,"prompt_tokens":1106,"completion_tokens":3054,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":722,"completion_tokens_details":{"reasoning_tokens":2978}},"tokens_in":722,"tokens_out":3054,"duration_ms":22155,"temperature":1.0,"reasoning_tokens":2978,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:32:27.657003+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct or find a signed bipartite dataset where the sign of edges is generated from high-rank structure, for example signs determined by per-node random vectors with no low-dimensional latent factor, then compare ELISE with its own variant that uses only the original-graph encoder: if the full model does not beat the variant, the low-rank assumption fails. The paper's own ablation tables provide the template for this test.","supporting_citations":[{"cited_title":"Balance in signed bipartite networks","cited_arxiv_id":null,"evidence_quote":"Defines balance in signed bipartite networks via signed butterflies, the balance-theoretic foundation that ELISE's message passing follows."},{"cited_title":"Personalized ranking in signed networks using signed random walk with restart","cited_arxiv_id":null,"evidence_quote":"Supplies the signed random walk with restart equations for score propagation that ELISE extends to node embeddings."},{"cited_title":"Lightgcn: Simplifying and powering graph convolution network for recommendation","cited_arxiv_id":null,"evidence_quote":"Provides the layer-wise aggregation and lightweight no-per-layer-weight design that ELISE adopts."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the randomized SVD algorithm used for the efficient low-rank refined graph."},{"cited_title":"Signed bipartite graph neural networks","cited_arxiv_id":null,"evidence_quote":"The signed-bipartite GNN baseline that adds same-type edges; ELISE must beat it and avoids its added edges."},{"cited_title":"Contrastive learning for signed bipartite graphs","cited_arxiv_id":null,"evidence_quote":"The contrastive signed-bipartite baseline; the main competitor for memory and speed on large graphs."},{"cited_title":"Deeper insights into graph convolutional networks for semi-supervised learning","cited_arxiv_id":null,"evidence_quote":"Identifies the over-smoothing failure mode that ELISE's personalized feature injection targets."}],"review_version":1}