{"id":"e628060a-6fe2-4b60-9395-b5ede5fdf533","arxiv_id":"2412.03097","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding initial residual connections and identity mapping to a LightGCN-style recommendation model yields small reported gains on Gowalla, Yelp-2018, and Amazon-Book.","lead":"This paper proposes a graph neural network recommendation model that adds residual connections and identity mapping to message-passing layers to reduce over-smoothing. It reports small accuracy gains over five baselines on three public datasets, but gives no code, hyperparameters, or uncertainty estimates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 0.0017–0.0033 Recall@20 gains over LGC are not load-bearing because the baseline identities and experimental conditions are unverified: GCMC and NGCF are referenced to unrelated papers, and LR-GCCF is named but omitted from Table 1.","rationale":"I read the paper in good faith: the proposed model is a plausible incremental combination of initial residual connections and identity mapping applied to GNN-based collaborative filtering, close in spirit to LR-GCCF and LightGCN. If the baseline results were verified and the small margins were reproducible, the paper could stand as a modest empirical contribution. However, the manuscript's own references and table undermine that empirical base. The reader's weakest assumption correctly identifies the validity of the baseline comparison as the load-bearing premise, and I agree with that diagnosis. The reference list mismatches for GCMC and NGCF, the missing LR-GCCF entry, and the absence of training details, code, and error bars mean the only quantitative evidence for the headline claim cannot be checked. Additionally, the abstract promises three models and interpretability tests that never appear, and no over-smoothing-specific measurement is provided; while these are secondary, they reinforce that the empirical case is incomplete. Because the central claim depends on a comparison that the paper does not substantiate, and because my concern is the same one the reader raised, I would keep the reader's REJECT verdict unchanged.","tokens_in":7247,"tokens_out":4533,"duration_ms":44913,"concrete_test":"Request the authors' code and exact configuration, then re-run Table 1 with the official public implementations of LightGCN, NGCF, GCMC, and LR-GCCF on the same 10-core preprocessed Gowalla, Yelp-2018, and Amazon-Book splits, fixing embedding size, layer count, learning rate, and the alpha/beta values used in Eq. (11), and reporting seed-averaged metrics. In particular, re-run LGC with identical hyperparameters; if the reproduced LGC Recall@20 on Gowalla differs from the reported 0.1754 by more than the claimed margin of 0.0017, or if LR-GCCF's Recall@20 exceeds 0.1771, the central claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section V.C, Table 1 is the only quantitative support for the central claim that the proposed algorithm outperforms four GNN-based baselines. For that comparison to be meaningful, the baselines must be the published algorithms and must have been run under the same protocol. This premise is not secure: reference [20], cited for GCMC, is a paper on tree-based feature selection for biological omics data, and reference [21], cited for NGCF, is a paper on temporal convolutional networks for action segmentation; neither is the corresponding recommender system. LR-GCCF is named in the text and in conclusion item 3, but it has no reference and no column in Table 1. With those citation failures, the table is consistent with comparisons against mislabeled or arbitrarily configured implementations. The reported improvements over LGC are tiny—about 0.0017–0.0033 in Recall@20 and about 0.0003–0.0033 in NDCG@20—so differences in data splits, embedding size, layer count, learning rate, regularization, or the values of alpha and beta could easily reverse or erase the margin. The paper supplies none of those training details, no code, and no error bars, making the comparison neither reproducible nor verifiable. What would have to be true for the central claim to hold is that the baseline results are accurate, correctly attributed, and obtained under fair conditions; the manuscript as written does not establish that.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a GNN-based collaborative filtering recommendation method that combines initial residual connections and identity mapping in the aggregation propagation layer, formalized in Eq. (11). The authors claim this design mitigates over-smoothing and improves recommendation accuracy, and they report recall and NDCG comparisons against BPRMF, GCMC, NGCF, LR-GCCF, and LGC on the Gowalla, Yelp-2018, and Amazon-Book datasets. The manuscript also mentions interpretability as a goal, although no explanation mechanism is described. The entire empirical case rests on a single results table with no error bars, no hyperparameters, no code, and no over-smoothing measurements.","tokens_in":7670,"tokens_out":5441,"duration_ms":52358,"significance":"If the claims were substantiated, the contribution would be an incremental variant of known residual/identity-mapping techniques applied to collaborative filtering recommenders, with marginal reported gains over a strong LightGCN baseline. The paper uses standard public datasets and addresses a real problem in GNN-based recommendation, but as submitted it provides no reproducible evidence: baseline identities are unsupported, the model is incompletely specified, and the over-smoothing mechanism is never directly tested. The significance of the work therefore cannot be assessed from the manuscript.","major_comments":[{"comment":"The baseline comparison is not verifiable because the citations and table do not match. Reference [20] is a tree-based feature-selection paper, not GCMC; reference [21] is a temporal convolutional network paper, not NGCF; reference [22] is the Neural Graph Collaborative Filtering paper but is cited for LR-GCCF; and the LR-GCCF column is absent from Table 1 even though the text and Conclusion item 3 name it. Consequently, the table does not establish that the proposed method outperforms the four named GNN baselines. The margins over LGC are small (e.g., 0.0003 to 0.0033 in NDCG@20), so even modest differences in data splits, hyperparameters, or baseline implementations could reverse the reported ordering.","section":"Section V.B and Table 1"},{"comment":"The paper's central mechanistic claim, that the propagation rule of Eq. (11) mitigates over-smoothing, is never tested. There is no depth sweep, no analysis of embedding similarity or rank collapse, and no ablation varying alpha, beta, or the number of layers. The assertion that an appropriate beta makes the maximum eigenvalue of (1-beta)I + beta W close to 1 and thereby reduces information loss is stated without derivation or measurement. Since over-smoothing is the paper's primary motivation, the experimental section provides no evidence for the proposed mechanism.","section":"Section IV.B and Section V.C"},{"comment":"The experimental setup is under-specified. The manuscript gives no data-splitting protocol, embedding size, number of propagation layers, learning rate, regularization coefficient, batch size, or values of alpha and beta. Table 1 reports a single set of numbers without error bars or standard deviations. These omissions make the claimed improvements non-reproducible and prevent assessment of whether the differences are statistically or practically significant.","section":"Section V"},{"comment":"The model is not fully specified. The layer-aggregation coefficients a_l are described as 'arbitrary real numbers,' but no training or selection procedure is given. The paragraph and equations are duplicated verbatim: Eq. (13) repeats Eq. (12), and Eq. (14) repeats Eq. (13). In addition, the abstract and introduction promise 'three' proposed models, but only a single model is described in Section IV, making the scope of the contribution unclear.","section":"Section IV.C, Eqs. (12)-(14)"},{"comment":"The paper repeatedly claims interpretability as a contribution, but the proposed method contains no explanation mechanism and no interpretability evaluation. The background discussion of explainable matrix factorization in Section III.A is never connected to the proposed GNN model. This claimed benefit is therefore unsupported and should either be demonstrated or removed from the claims.","section":"Abstract and Section II"}],"minor_comments":[{"comment":"The formulas contain corrupted or placeholder symbols, such as undefined subscripts in Eqs. (1) and (2), and are not introduced coherently. The figure caption 'Second Order Phase Diagram and Rating Matrix' also appears to be a template artifact unrelated to the content.","section":"Section III.A, Eqs. (1)-(5)"},{"comment":"Several references do not support the sentences they are attached to. For example, [14] is about exposure modeling rather than residual weights, and [17] is about survival prediction rather than eigenvalue analysis. The authors should cite the actual sources for residual connections, identity mapping, and over-smoothing or remove the mismatched citations.","section":"References [3]-[18]"},{"comment":"The duplicated paragraph and equations (Eqs. (12)-(14)) should be consolidated and the equation numbering unified, as the current text repeats the same derivation twice.","section":"Section IV.C"},{"comment":"Reference [22] is labeled LR-GCCF in the text but is actually the NGCF paper. The authors should either add the correct LR-GCCF reference (He et al., SIGIR 2020) or remove the baseline name from the comparison.","section":"Section V.B and References"}],"recommendation":"reject","confidential_remarks":"The manuscript has multiple citation-to-claim mismatches beyond the baselines, including unrelated references for residual connections and over-smoothing. In addition to the technical problems detailed in the report, the pattern of unrelated citations is a scholarly-integrity concern that the editor may wish to examine. The paper appears to be an unpolished excerpt rather than a complete, self-contained submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me skip straight to it: this manuscript presents a reasonable incremental tweak to LightGCN—initial residual connections plus identity mapping in the aggregation step—but the experimental section is not credible as written.\n\nWhat is genuinely new is narrow. Equation (11) is a specific way of combining (1-alpha)S H(l-1) + alpha H(0) with (1-beta)I + beta W(l-1). That's APPNP-style initial residuals mated with ResNet-style identity mapping, both published and cited. There's no theory, no depth sweep, no embedding-similarity analysis, so the over-smoothing claim is only assumed.\n\nTo its credit, the model is easy to state, and the idea that these two known tricks work together is plausible. The gains reported are small but consistent across three datasets, which at least suggests the variant isn't hurting.\n\nThe soft spots are big. The worst is the baseline list. Reference [20] is a paper on tree-based feature selection for omics data, not GCMC; reference [21] is a TCN paper on action segmentation, not NGCF. LR-GCCF is named in the text but missing from Table 1 and from the reference list. That means the table, which is the entire empirical support, could easily be comparing against whatever implementations the authors happened to run—there is no way to verify. No hyperparameters, no code, no error bars. The margins over LGC are 0.0017–0.0033 in Recall@20; that's noise territory without matched settings.\n\nAlso, the abstract promises three models and interpretability tests that never show up. The eigenvalue remark about beta is asserted without proof. These aren't minor—they compound the basic reproducibility problem.\n\nI think the right call is desk reject. The method is not important enough to referee with the current evidence. If the authors come back with the correct baselines, full training details, code, and an actual over-smoothing measurement, then it might be a small but legitimate contribution.\n\nSo: no to peer review, would not cite, don't bring to reading group unless you want to teach citation integrity.","headline":"A plausible but unsupported LightGCN tweak; the baseline citations don't match the named methods, so the empirical claim is unattested.","tokens_in":8096,"tokens_out":2877,"would_cite":false,"duration_ms":26380,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that adding initial residual connections and identity mapping to a GNN collaborative-filtering propagation layer prevents over-smoothing and yields small consistent Recall@20 and NDCG@20 gains over five baselines on three…","keywords":["Recommendation Systems","Graph Neural Networks","Over-Smoothing","Collaborative Filtering","Interpretability","Initial Residual Connections","Identity Mapping"],"falsifier":"Run the same comparison on Gowalla, Yelp-2018, and Amazon-Book with embedding size, number of layers, learning rate, $\\alpha$, $\\beta$, and regularization reported, include LR-GCCF in the table, and evaluate across multiple seeds; if LGC or any full baseline matches or exceeds the proposed model under matched tuning, the paper's central claim is refuted.","tokens_in":7029,"feed_emoji":"📈","tokens_out":9700,"duration_ms":79872,"temperature":0.7,"pith_summary":"Over-smoothing is the failure mode where deep graph convolutions drive all node embeddings toward the same vector, so adding layers stops helping. This paper proposes a GNN-based collaborative filtering model whose propagation rule keeps the initial user and item embeddings alive at every depth through initial residual connections, and mixes in a trainable identity-mapped transformation whose eigenvalues are tuned to reduce information loss. The paper claims this design outperforms BPRMF, GCMC, NGCF, LR-GCCF, and LGC in Recall@20 and NDCG@20 on Gowalla, Yelp-2018, and Amazon-Book. The reported gains are small but consistent, for example Recall@20 of 0.1771 versus 0.1754 for LGC on Gowalla. A sympathetic reader should care because the recipe is simple: a one-line change to the propagation layer that may allow recommender networks to go deeper without collapsing.","feed_headline":"Residual links stop GNN over-smoothing and lift recommendation recall","feed_subtitle":"A GNN model with initial residuals and identity mapping edges out five baselines on three datasets.","key_machinery":"The engine is the aggregation-propagation update $\\mathbf{H}^{(l)} = \\sigma(((1-\\alpha)\\mathbf{S}\\mathbf{H}^{(l-1)} + \\alpha\\mathbf{H}^{(0)})((1-\\beta)\\mathbf{I} + \\beta\\mathbf{W}^{(l-1)}))$, where $\\mathbf{S}$ is the normalized adjacency (Laplacian) matrix of the user-item bipartite graph, $\\mathbf{H}^{(0)}$ is the initial embedding matrix, $\\sigma$ is a nonlinear activation, and $\\mathbf{W}^{(l-1)}$ is trainable. Two scalar knobs carry the argument: $\\alpha$ sets how much of the first-layer embedding survives into every later layer (initial residual connection), and $\\beta$ controls the identity mapping so that the eigenvalues of $(1-\\beta)\\mathbf{I} + \\beta\\mathbf{W}^{(l-1)}$ stay close to 1, which the paper says limits information loss during propagation. Layer outputs are then combined as $\\mathbf{H}^{(*)} = \\sum_{l=0}^{L} a_l \\mathbf{H}^{(l)}$ with free weights $a_l$, and the predicted rating for a user-item pair is the inner product of the aggregated user and item embeddings. This combination is what lets the model pool collaborative signals from different propagation depths instead of relying on the last layer alone.","core_discovery":"The central claim, stated on the paper's own terms, is that the over-smoothing problem in graph-based recommendation can be solved at the propagation layer without redesigning the graph or the loss. The layer update $\\mathbf{H}^{(l)} = \\sigma(((1-\\alpha)\\mathbf{S}\\mathbf{H}^{(l-1)} + \\alpha\\mathbf{H}^{(0)})((1-\\beta)\\mathbf{I} + \\beta\\mathbf{W}^{(l-1)}))$ preserves the initial embedding $\\mathbf{H}^{(0)}$ at every depth and mixes it with a normalized adjacency propagation $\\mathbf{S}\\mathbf{H}^{(l-1)}$, while the term $(1-\\beta)\\mathbf{I} + \\beta\\mathbf{W}^{(l-1)}$ gives the network a nonlinear, trainable transformation whose largest eigenvalue can be held near 1. The authors report that this rule beats the matrix-factorization baseline BPRMF and four GNN baselines on the Gowalla, Yelp-2018, and Amazon-Book datasets in both Recall@20 and NDCG@20, and attribute the improvement to retaining initial information and avoiding convergence to a subspace. The paper also frames the design as a step toward interpretable recommendations, since the layered aggregation makes the contribution of each propagation depth visible.","pith_inferences":["Our inference: the same propagation layer should transfer to other bipartite-graph prediction tasks, such as link prediction or friend recommendation, because nothing in the mechanism depends on ratings being the prediction target.","Our inference: the reported gains are small enough that seed variance and statistical significance matter; a production system should not be switched on the strength of a 0.002-0.003 Recall difference without replication.","Our inference: because the paper does not report $\\alpha$, $\\beta$, layer count, embedding size, or regularization, the natural next experiment is an ablation sweep over those knobs to see whether the initial residual or the identity mapping carries the gain.","Our inference: the eigenvalue argument suggests that even deeper stacks than the ones tested here could remain useful, but that extrapolation is untested in the paper."],"forward_implications":["If the paper is right, recommender backbones can be made deeper without the usual accuracy cliff, because the initial embeddings are never washed out by repeated convolutions.","Long-range collaborative signals from higher-order neighbors can be combined with first-hop evidence at prediction time, since every layer's output enters the final weighted sum.","The reported numbers give a concrete target for later anti-over-smoothing recommenders: on Gowalla, Recall@20 of 0.1771 and NDCG@20 of 0.1491.","The design suggests that tuning $\\beta$ near the boundary where the identity-mapping eigenvalues approach 1 is a practical lever for trading information retention against nonlinear feature extraction."],"supporting_citations":[{"why":"Supplies BPRMF, the matrix-factorization baseline that all GNN models are required to beat.","marker":"[19]"},{"why":"Marker under which the paper cites the GCMC baseline in the comparison table.","marker":"[20]"},{"why":"Marker under which the paper cites the NGCF baseline.","marker":"[21]"},{"why":"Marker under which the paper cites the LR-GCCF baseline name in the text.","marker":"[22]"},{"why":"Supplies LightGCN (LGC), the strongest comparator in the reported table.","marker":"[23]"},{"why":"Supports the need for initial residual connections by arguing that GNNs lose expressive power in deep layers.","marker":"[13]"},{"why":"Supports introducing the scalar $\\alpha$ to balance previous-layer propagation against initial residuals.","marker":"[14]"},{"why":"Supports the beta-controlled identity mapping by linking its eigenvalues to information loss during propagation.","marker":"[17]"}],"fun_headline_variants":["Residual GNN propagation preserves user data, lifts recall","Identity mapping in GNNs halts over-smoothing","New GNN layer beats matrix factorization and 4 baselines","Over-smoothing defeated by residual initial embeddings","GNN recommendation: keep starting info to top-20 hits"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim collapses if the named baselines were not implemented as their published algorithms and tuned under matched conditions, since the paper gives no training details, the table omits the LR-GCCF column it claims to compare, and several baseline citations do not point to the cited algorithms.","fun_headline_variants_meta":{"raw":{"variants":["Residual GNN propagation preserves user data, lifts recall","Identity mapping in GNNs halts over-smoothing","New GNN layer beats matrix factorization and 4 baselines","Over-smoothing defeated by residual initial embeddings","GNN recommendation: keep starting info to top-20 hits"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000792,"raw_usage":{"total_tokens":3503,"prompt_tokens":970,"completion_tokens":2533,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":586,"completion_tokens_details":{"reasoning_tokens":2451}},"tokens_in":586,"tokens_out":2533,"duration_ms":17979,"temperature":1.0,"reasoning_tokens":2451,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:46:34.152200+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same comparison on Gowalla, Yelp-2018, and Amazon-Book with embedding size, number of layers, learning rate, $\\alpha$, $\\beta$, and regularization reported, include LR-GCCF in the table, and evaluate across multiple seeds; if LGC or any full baseline matches or exceeds the proposed model under matched tuning, the paper's central claim is refuted.","supporting_citations":[{"cited_title":"Comparison of Tree-Based Feature Selection Algorithms on Biological Omics Dataset,","cited_arxiv_id":null,"evidence_quote":"Marker under which the paper cites the GCMC baseline in the comparison table."},{"cited_title":"Adaptive ReceptiveField U-ShapedTemporalConvolutionalNetworkfor Vulgar ActionSegmentation,","cited_arxiv_id":null,"evidence_quote":"Marker under which the paper cites the NGCF baseline."},{"cited_title":"Neural graph collaborative filtering","cited_arxiv_id":null,"evidence_quote":"Marker under which the paper cites the LR-GCCF baseline name in the text."},{"cited_title":"LightGCN: Simplifying and powering graph convolution network for recommendation","cited_arxiv_id":null,"evidence_quote":"Supplies LightGCN (LGC), the strongest comparator in the reported table."},{"cited_title":"Modeling user exposure in recommendation","cited_arxiv_id":null,"evidence_quote":"Supports introducing the scalar $\\alpha$ to balance previous-layer propagation against initial residuals."},{"cited_title":"Survival prediction across diverse cancer types using neural networks,","cited_arxiv_id":null,"evidence_quote":"Supports the beta-controlled identity mapping by linking its eigenvalues to information loss during propagation."}],"review_version":1}