{"id":"856cc7d0-6b19-4fc4-b841-b6262b437075","arxiv_id":"2509.05115","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"A recommender that uses MF and SVD to build contrastive views claims improved recall/NDCG on small datasets, but its core fusion equations are ill-defined and its experimental numbers don't consistently match.","lead":"The paper proposes HMFGCL, a graph contrastive learning recommender that builds contrastive views from both low-rank matrix factorization and SVD of the user-item adjacency matrix, plus a mixed-noise GNN encoder. It reports 5-11% relative gains over baselines on three small user-item datasets, but the key view-fusion equations appear dimensionally invalid and the reported statistics are internally inconsistent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eqs. 16-21 define the MF-SVD contrastive-view fusion with incompatible matrix dimensions, so the central mechanism is unimplementable as written.","rationale":"Reading in good faith, the paper's core claim is that hybrid MF-SVD view fusion, plus mixed noise, yields consistent gains. That claim would require the fused embeddings G^(u) and G^(v) in Eqs. 16-21 to be well-defined d-dimensional vectors. They are not: the matrix products are dimensionally invalid (Ã_MF E Â_SVD has an undefined last product; the per-user version collapses to a scalar if corrected). The InfoNCE losses in Eqs. 22-23 depend on these embeddings, so the entire method cannot be instantiated as written. This is an internal mathematical defect, not a disagreement with the field's consensus, and it is not rescued by Algorithm 1's high-level 'fusion' calls. Independent support such as released code or machine-checked proof is absent, and the reported experimental inconsistencies (e.g., Table 1 vs Table 2 values for ML-100K) further reduce confidence. Since the reader already rejected the paper on this basis, our pass leaves the verdict unchanged.","tokens_in":18623,"tokens_out":6042,"duration_ms":60514,"concrete_test":"Use the ML-100k shapes (m=943, n=1682, d=64) and execute Eq. 17 literally in PyTorch: the final torch.matmul between an m×d tensor and an m×q matrix raises a dimension-mismatch error, confirming the formulas cannot be run. Then ask the authors to provide a version of Eq. 20 that produces a d-vector from the same factors; if no correction yields both G and g with consistent dimensions, the view-fusion step is underspecified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. 16 defines the fused user view as G_l^(u) = sigma(Ã_MF E_{l-1}^(v) Â_SVD), with Ã_MF in R^{m×n}, E_{l-1}^(v) in R^{n×d}, and Â_SVD in R^{m×n}. The product Ã_MF E_{l-1}^(v) is m×d, and multiplying it by Â_SVD (m×n) is undefined. Expanding via Eq. 17 does not repair this: ˜P_k ˜Q_k^T E^{(v)}_{l-1} ˆP_q ˆS_q ˆQ_q^T is m×d, then multiplied by ˆP_q (m×q) fails. The per-user version in Eq. 20, g_{m,l}^{(u)} = sigma(Ã_{m,:} E^{(v)}_{l-1} Â_{m,:}), has the same defect: a 1×n row times an n×d matrix gives 1×d, which cannot be multiplied by another 1×n row; transposing the last factor gives a scalar, not a d-dimensional embedding. Because G_l^(u), G_l^(v), g_{m,l}^{(u)}, and g_{n,l}^{(v)} are the contrastive views that enter InfoNCE (Eqs. 22-23), the entire training signal is defined only if the reader silently replaces the printed formulas with some unstated variant. No code or formal proof is provided to resolve this ambiguity. This is an internal mathematical inconsistency in the core construction, not a disagreement with existing baselines; it makes the claimed 7-11% improvements in Sec. 5.3 unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HMFGCL, a graph contrastive learning recommender that fuses two low-rank matrix factorizations (MF and SVD) of the user-item adjacency matrix to construct contrastive views, and adds mixed noise during GNN propagation. The claimed contribution is that the hybrid MF-SVD views capture complementary global collaborative information, leading to improvements over several baselines, especially on small datasets. Experiments are reported on ML-100k, ModCloth-1000, and LastFM-1000 against twelve baselines, with relative gains over LightGCL of roughly 7–11% claimed in Section 5.3. The paper includes an algorithm listing, hyperparameter sensitivity studies, loss convergence plots, and an ablation study.","tokens_in":19066,"tokens_out":4713,"duration_ms":50723,"significance":"If the proposed method were correctly specified and the results reproducible, the paper would offer a modest incremental contribution to graph contrastive learning for recommendation: using complementary low-rank factorizations to build contrastive views, combined with mixed-noise feature perturbation, is a plausible recipe. However, the central fusion formulas are dimensionally invalid, so the method as written cannot be implemented, and the experimental reporting contains internal numerical inconsistencies. The paper also ships no code and provides no machine-checkable derivations. Thus the significance is conditional and cannot be assessed from the present manuscript.","major_comments":[{"comment":"The central contrastive-view construction is dimensionally invalid. With \\tilde A_MF ∈ R^{m×n}, E^{(v)}_{l-1} ∈ R^{n×d}, and \\hat A_SVD ∈ R^{m×n}, the product \\tilde A_MF E^{(v)}_{l-1} is m×d and cannot be right-multiplied by \\hat A_SVD, which is m×n. The expanded form in Eq. (17) does not repair this: \\tilde P_k \\tilde Q_k^T E^{(v)}_{l-1} \\hat P_q \\hat S_q \\hat Q_q^T yields an m×d matrix that cannot then be multiplied by \\hat P_q (m×q). The per-user version in Eq. (20) has the same defect: a 1×n row times an n×d matrix gives 1×d, which cannot be multiplied by the 1×n row \\hat A_{m,:}; transposing the last factor yields a scalar, not a d-dimensional embedding. Since G_l^(u), G_l^(v), g_{m,l}^{(u)}, and g_{n,l}^{(v)} are the contrastive views that enter the InfoNCE losses in Eqs. (22)–(23), the entire training signal is undefined unless the reader silently replaces the printed formulas wi","section":"§4.4, Eqs. (16)–(21)"},{"comment":"The reported performance claims are inconsistent with the tables. The text states that on LastFM-1000, R@10 improved by 10.24% over LightGCL, but from Table 1 the improvement is (0.1815−0.1679)/0.1679 = 8.10%; the other three LastFM percentages (N@10 7.85%, R@20 7.39%, N@20 7.44%) match the table. Additionally, the HMFGCL row for ML-100k reports N@20=0.4001 in Table 1, whereas Table 2 and §5.4.1 report N@20=0.4008 for the selected dimension; R@20 differs as well (0.334 vs 0.3351). These numerical discrepancies undermine the headline comparison and prevent verification of the claimed gains.","section":"§5.3, Table 1; §5.4.1, Table 2"},{"comment":"The experimental protocol does not support the claimed 7–11% improvements. No validation split is described: the hyperparameter studies in §5.4.1–§5.4.4 select dimensions, singular values, layer counts, and embedding sizes by test-set performance (e.g., dimension 5 is chosen because it maximizes test R@10/N@10). No error bars, multiple-seed runs, or significance tests are reported. Since the baselines are taken from RecBole with 'optimal parameters' while HMFGCL is tuned on the test sets, the reported margins may reflect tuning advantage rather than a robust algorithmic improvement.","section":"§5.3 and §5.4"}],"minor_comments":[{"comment":"The regularization term in Eq. (2) is typeset as 'λ E0 2'; it should presumably be λ‖E0‖² or similar. As printed, the term is unreadable.","section":"Eq. (2)"},{"comment":"Line 18 returns 'ˆPq, ˆQq, ˆQq'; the third return should be ˆSq (the singular values). Line 20 calls GNN(E^(u), A) for the item view; it should use E^(v).","section":"Algorithm 1, lines 18 and 20"},{"comment":"The text refers to the 'LastMF-1000' dataset; this should be 'LastFM-1000'. The caption of Fig. 17 uses the correct name.","section":"§5.5"},{"comment":"The caption says the best results are in bold and second-best underlined, while the text says best results are highlighted with gray shading; the table as rendered uses asterisks. This formatting description is inconsistent and should be corrected.","section":"Table 1 caption and §5.3 text"},{"comment":"The item aggregation z^(v)_{n,l} = σ(A_{:,n}·E^(u)'_{l-1}) appears dimensionally problematic: A_{:,n} is an m×1 column, while E^(u)' is m×d. A transposed row or inner-product formulation is likely intended.","section":"Eq. (15)"}],"recommendation":"reject","confidential_remarks":"The stress-test concern about Eqs. (16)–(21) lands directly: the dimension mismatch is an internal inconsistency in the core construction, not a disagreement with an external benchmark. The numerical discrepancies in Section 5.3 further compound the problem. In my view, the manuscript would require a complete rewrite of the method specification, a corrected experimental protocol with validation-based hyperparameter selection and variance reporting, and a public code release before it could be meaningfully reviewed. I see no path to acceptance from the current text."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper's central equations are dimensionally wrong, so HMFGCL cannot be implemented as described, and the experimental reporting has inconsistencies that undermine the headline numbers. The idea itself isn't crazy—combining MF and SVD for contrastive view generation is a natural extension of LightGCL—and the authors state the motivation clearly: random edge dropout discards information, while SVD-guided views capture global structure. The ablation suggests both factorization branches contribute, which is a point in their favor.\n\nThe problem is that Section 4.4 is the heart of the method and it doesn't type-check. Eq. 16 defines G_l^(u) = σ(Ã_MF E^(v)_{l-1} Â_SVD). With Ã_MF an m×n matrix, E^(v) n×d, and Â_SVD m×n, the product can't be formed. Expanding to Eq. 17 doesn't help—after the first multiplication you have an m×d matrix, and multiplying by Â_SVD (m×n) is undefined. The per-user variant in Eq. 20 has the same defect. So the contrastive views that feed the InfoNCE loss (Eqs. 22–23) are undefined. There is no code or formal proof to fall back on. This is a load-bearing flaw, not a typo.\n\nThe experimental section is also shaky. Section 5.3 claims a 10.24% R@10 improvement over LightGCL on LastFM-1000, but the numbers in Table 1 imply 8.10%. HMFGCL's ML-100K R@20 and N@20 in Table 1 (0.334, 0.4001) disagree with Table 2 (0.3351, 0.4008). No error bars, no validation split, and hyperparameters appear to be tuned on the test set, so the reported gains are not robust evidence.\n\nReaders who work on graph contrastive learning for collaborative filtering might still get a useful hint here—the hybrid factorization view is worth exploring. But this particular version would need corrected equations, a working implementation, and a cleaner experimental protocol before it deserves referee time. My advice: desk reject and let the authors resubmit a repaired version.","headline":"The core fusion equations in Section 4.4 are dimensionally invalid, so HMFGCL as written cannot be implemented; the experimental numbers also have internal inconsistencies. The underlying idea is a plausible extension of LightGCL, but this version needs major repair.","tokens_in":19522,"tokens_out":6191,"would_cite":false,"duration_ms":56858,"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":"HMFGCL fuses matrix factorization with SVD to build contrastive views for graph recommendation, reporting consistent gains over LightGCL on three datasets.","keywords":["Graph contrastive learning","Recommendation system","Matrix factorization","Singular value decomposition","Data augmentation","Collaborative filtering","GNN","Noise injection"],"falsifier":"Implement Eq. 16 on a dataset with m ≠ n: the product Ã_MF E(v)_{l-1} Â_SVD is undefined because E(v) is n×d while Â_SVD is m×n, so the forward pass cannot run as written. A corrected construction that produces well-defined embeddings and reproduces the reported 7–11% gains would settle the claim; alternatively, the same experiments with the fusion order changed to a valid product would show whether the gains come from the hybrid views or from other components.","tokens_in":18504,"feed_emoji":"📊","tokens_out":6749,"duration_ms":65105,"temperature":0.7,"pith_summary":"The paper argues that the two standard ways of generating contrastive views in graph-based recommendation—randomly perturbing the user–item graph and clustering nodes—do not extract the full interaction signal. It proposes HMFGCL, which instead builds contrastive views by fusing two low-rank factorizations of the adjacency matrix: ordinary matrix factorization (MF) and singular value decomposition (SVD). A graph neural network generates the main view, mixed Gaussian and uniform noise is injected during propagation, and the MF/SVD fusion supplies the augmented view for an InfoNCE contrastive loss. Across MovieLens-100k and 1000-user subsets of ModCloth and LastFM, the authors report relative improvements of roughly 7–11% over the best baseline, LightGCL, with ablations showing both factorizations contribute. If the claim holds, it means a cheap preprocessing-level augmentation can improve recommendation accuracy on sparse, small graphs without random edge loss.","feed_headline":"MF+SVD contrastive views lift graph recommenders by up to 11%","feed_subtitle":"Hybrid MF and SVD factorization builds views that beat LightGCL on three small recommendation datasets.","key_machinery":"The load-bearing mechanism is the fusion equation G(u)_l = σ(Ã_MF E(v)_{l-1} Â_SVD), which combines the MF-reconstructed adjacency, the previous layer's item embeddings, and the SVD-reconstructed adjacency to form a user contrastive view; a transposed variant forms item views. The second mechanism is mixed noise injection, Δ = ω ⊙ (l1 N_g + l2 N_u), added to embeddings before each aggregation to improve representation uniformity. The fusion is meant to inject global collaborative structure into each layer's local message passing, while the noise is meant to regularize the embeddings before contrastive comparison.","core_discovery":"The central claim is that complementary low-rank decompositions capture global collaborative information that local message passing and single-view augmentations miss. Concretely, HMFGCL factorizes the normalized adjacency matrix A into an MF approximation Ã_MF = P̃_k Q̃_k^T and an SVD approximation Â_SVD = P̂_q Ŝ_q Q̂_q^T, then fuses these with the layer-wise GNN embeddings to form contrastive views for users and items. The main view is the standard GNN output; the contrastive view is the global-information-fused embedding, and the two are compared with a layer-wise InfoNCE loss. The paper reports that this configuration outperforms all baselines on three datasets, and that removing either","pith_inferences":["A natural test of the complementarity story is to swap in another low-rank construction (e.g., nonnegative MF or CUR) and see whether the gains persist; if they do, the specific pair matters less than having two differently regularized views.","Because the preprocessing is graph-agnostic, the same hybrid could be applied to item-side or session graphs, not just bipartite user–item graphs.","The reported results are on very small graphs; a scaled-up study on full-size datasets would clarify whether the gains survive when SVD computation becomes expensive.","If the fusion step is repaired to be dimensionally coherent, the method could be compared with LightGCL under identical training budgets to separate the effect of the view construction from the noise injection."],"forward_implications":["Random edge/node dropout becomes unnecessary for view generation, so sparse graphs avoid further information loss.","Small graphs benefit most: the largest margins appear on the 1000-user datasets, suggesting low-rank global structure is especially valuable when local neighborhoods are sparse.","The two factorizations contribute non-redundantly: each singleton-factorization variant beats the no-factorization baseline, and the combination beats both.","Two GNN layers suffice for the reported gains, so the added cost is mostly the one-time MF/SVD preprocessing rather than deeper networks."],"supporting_citations":[{"why":"Supplies the SVD-guided graph augmentation idea and the strongest baseline that HMFGCL extends by adding MF as a complementary view source.","marker":"[11]"},{"why":"Introduces noise-based contrastive views without structural augmentation, the mechanism HMFGCL modifies into mixed Gaussian and uniform noise.","marker":"[13]"},{"why":"Provides the random graph augmentation baseline (SGL) and motivates the claim that edge-drop views lose information on sparse graphs.","marker":"[9]"},{"why":"Supplies the LightGCN-style GNN backbone used for neighborhood aggregation and the main-view embeddings.","marker":"[30]"},{"why":"Provides XSimGCL, the cross-layer contrastive baseline that the paper compares against and builds on.","marker":"[14]"},{"why":"Represents the clustering-based augmentation strategy that the paper argues fails to capture full user-item interactions.","marker":"[12]"},{"why":"Supplies the open-source recommendation framework used to run all baseline implementations and hyperparameter settings.","marker":"[33]"}],"fun_headline_variants":["Hybrid MF+SVD views sharpen graph contrastive recommenders","Graph contrastive learning gets a boost from hybrid factorization","Two-factor views beat single-view graph recommenders","MF+SVD fusion improves graph contrastive recommendation","Complementary decompositions refine graph contrastive views"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The fusion step that creates the contrastive views assumes the matrix products in Equations 16–21 are well-defined with the stated shapes; on a non-square user–item matrix those products do not multiply as written, so the reported results depend on that step being coherent.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid MF+SVD views sharpen graph contrastive recommenders","Graph contrastive learning gets a boost from hybrid factorization","Two-factor views beat single-view graph recommenders","MF+SVD fusion improves graph contrastive recommendation","Complementary decompositions refine graph contrastive views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1090,"prompt_tokens":727,"completion_tokens":363,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":471,"completion_tokens_details":{"reasoning_tokens":298}},"tokens_in":471,"tokens_out":363,"duration_ms":4411,"temperature":1.0,"reasoning_tokens":298,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:35:29.138139+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement Eq. 16 on a dataset with m ≠ n: the product Ã_MF E(v)_{l-1} Â_SVD is undefined because E(v) is n×d while Â_SVD is m×n, so the forward pass cannot run as written. A corrected construction that produces well-defined embeddings and reproduces the reported 7–11% gains would settle the claim; alternatively, the same experiments with the fusion order changed to a valid product would show whether the gains come from the hybrid views or from other components.","supporting_citations":[{"cited_title":"Are graph augmen- tations necessary? simple graph contrastive learning for recommendation","cited_arxiv_id":null,"evidence_quote":"Introduces noise-based contrastive views without structural augmentation, the mechanism HMFGCL modifies into mixed Gaussian and uniform noise."},{"cited_title":"Self-supervised graph learning for recommendation","cited_arxiv_id":null,"evidence_quote":"Provides the random graph augmentation baseline (SGL) and motivates the claim that edge-drop views lose information on sparse graphs."},{"cited_title":"Lightgcn: Simplifying and powering graph convolution network for recommen- dation","cited_arxiv_id":null,"evidence_quote":"Supplies the LightGCN-style GNN backbone used for neighborhood aggregation and the main-view embeddings."},{"cited_title":"Xsimgcl: Towards extremely simple graph contrastive learning for recom- mendation.IEEE Transactions on Knowledge and Data Engineering, 2023","cited_arxiv_id":null,"evidence_quote":"Provides XSimGCL, the cross-layer contrastive baseline that the paper compares against and builds on."},{"cited_title":"Improving graph collaborative filtering with neighborhood-enriched contrastive learning","cited_arxiv_id":null,"evidence_quote":"Represents the clustering-based augmentation strategy that the paper argues fails to capture full user-item interactions."},{"cited_title":"Recbole: Towards a unified, comprehensive and efficient framework for recommenda- tion algorithms","cited_arxiv_id":null,"evidence_quote":"Supplies the open-source recommendation framework used to run all baseline implementations and hyperparameter settings."}],"review_version":1}