{"id":"e71235ff-6b82-4139-a741-2ed1e3c6767b","arxiv_id":"2506.02834","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A recommender that sums interaction, social-friendship, and quantized Jaccard-similarity graph signals in a LightGCN-style model edges out existing social recommenders on four benchmarks.","lead":"This paper combines users' social friendships, shopping-behavior similarity, and past purchases into one graph-convolution recommender system. It reports small but consistent accuracy gains over several existing recommendation models on four public datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The user-correlation matrix C defined by equations (4)-(6) is internally inconsistent: WU ⊙ WI^{-1} equals |Ii ∪ Ij|/|Ii ∩ Ij| ≥ 1, while Table 2 bins only [0,1]; C is therefore unspecified as written.","rationale":"The paper's contribution over LightGCN and WiGCN is the addition of the user-correlation matrix C, built from quantized Jaccard similarities and propagated alongside A and S in equations (9)-(14). The reader's weakest assumption correctly identifies the potential redundancy and the concrete formula error in equation (6). My stress-test agrees that the formula error is the single most load-bearing issue because it prevents even an attempt at reproduction: the mapping in Table 2 is defined only for [0,1], while the printed expression yields values ≥1. The text at (6) additionally contradicts equation (5), making the intended definition ambiguous. This is not a mere stylistic flaw; the ablation results in Tables 5, 7, and 8 all attribute gains to C, so without a corrected definition the central empirical claim is unverifiable. The conditional verdict is appropriate: the paper should be revised to correct equations (4)-(6), release code/data, and report variance. My read does not change the verdict; it sharpens the specific blocker.","tokens_in":16851,"tokens_out":6044,"duration_ms":63307,"concrete_test":"Implement GCCF twice on Gowalla and LibraryThing with identical hyperparameters, differing only in C: (a) C = f(WI ⊙ WU^{-1}) (Jaccard = |Ii ∩ Ij| / |Ii ∪ Ij|), and (b) C = f(WU ⊙ WI^{-1}) with values >1 discarded or clamped to the nearest bin. Compare recall@20, precision@20 and NDCG@20 to Table 4 'Our w/ interact' and 'Our model-all'. If only (a) reproduces the reported numbers, the typo is confirmed and the model is implementable; if neither matches, the central results are not supported by the written method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation (4) sets WU = |Ii ∪ Ij| and equation (5) sets WI = |Ii ∩ Ij|. Equation (6) then defines C = f(WU ⊙ WI^{-1}). Element-wise this is |Ii ∪ Ij| / |Ii ∩ Ij|, which is ≥ 1 whenever the two item sets overlap and is undefined when they are disjoint. Yet Table 2 classifies only values in [0,1) as Jaccard indices, and the text at (6) reverses the roles, calling WI the union matrix. Thus, the central new signal in GCCF — the quantized user-correlation matrix C — cannot be constructed from the paper as written. If the literal formula is used, C contains no entries in the tabled bins (or requires an unspecified clamping), making 'Our w/ interact' and 'Our model-all' equivalent to a model without C; if the intended formula was WI ⊙ WU^{-1}, the paper still must be corrected. Because every reported improvement over LightGCN/WiGCN in Tables 4, 7 and 8 is attributed to this matrix, the empirical central claim is not reproducible from the specification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript presents GCCF, a graph convolution collaborative filtering model that augments a LightGCN-style user-item propagation with two additional user-user signals: a quantized Jaccard overlap matrix C computed from purchase/rating histories and a social friendship matrix S. The authors propose a preprocessing pipeline (10-core filtering and user retention by Jaccard overlap), a propagation layer that averages normalized signals from A, C, and S, and a BPR loss. They compare against nine baselines on Gowalla, Librarything, Ciao, and Epinions and report that the full model achieves the highest recall, precision, and NDCG@20, while also converging faster than LightGCN-based baselines.","tokens_in":17199,"tokens_out":9391,"duration_ms":96302,"significance":"If the empirical claims are correct, the contribution is modest but useful: it shows that a lightweight, transformation-free GCN can incorporate social and behavioral user-user correlation signals, and it provides an ablation separating the two signals. The paper compares against a broad set of baselines and reports per-dataset results, which is a strength. However, the central new signal C is not well defined as written, the reported gains are small and without variance estimates, and the preprocessing makes the evaluation non-standard. Because of these issues, the significance of the claims cannot be assessed from the manuscript in its current form.","major_comments":[{"comment":"Equations (4) and (5) define WU = |Ii ∪ Ij| and WI = |Ii ∩ Ij|. Equation (6) then sets C = f(WU ⊙ WI^{-1}), whose element is |union| / |intersection|, which is ≥ 1 whenever the intersection is nonempty and undefined when it is empty. Table 2, however, partitions only values in [0,1] as Jaccard indices, and the sentence after (6) calls WI 'the matrix that represents the union.' The literal reading cannot produce the C used in the experiments; if the intended formula is WI ⊙ WU^{-1}, the paper must state this, handle the zero-intersection case, and re-run the experiments, because every reported improvement over LightGCN/WiGCN is attributed to this matrix.","section":"III-A-2, Eqs. (4)-(6), Table 2"},{"comment":"The manuscript does not state whether the correlation matrix C is constructed from the full preprocessed interaction matrix or only from the 80% training split used for learning. If C is built from all interactions before the 20% latest-interaction test split, then test labels are visible in the user-user overlap signal during training, which would invalidate the evaluation. The authors should specify that C, S, and the propagation graph are computed from training data only, and recompute all tables if necessary.","section":"III-A-2 and IV-A"},{"comment":"The text in Section IV-B states that 'our model with interaction embedding (Our w/ interaction) still gives the highest results' among the non-social models MF, GCMC, NGCF, LightGCN, and WiGCN. Table 4 contradicts this on Epinions: for recall, LightGCN achieves 0.0150 versus 0.0139 for 'Our w/ interact', and for NDCG@20, 0.0085 versus 0.0073. This claim needs to be corrected or the table corrected.","section":"IV-B, Table 4"},{"comment":"The quantization function f in Table 2 uses hand-selected bin boundaries and typical values (0.0, 0.005, 0.05, 0.5, 1.0) with no stated selection procedure and no sensitivity analysis. Because the reported improvements of the full model over LightGCN are small (e.g., +7.3% recall on Gowalla, +4.7% on Librarything; interaction-only gains of +2.1% and +0.3%), it is not possible to tell whether the gains come from the proposed propagation or from threshold choices tuned to the datasets. The authors should justify the thresholds via validation and report sensitivity or ablation results.","section":"III-A-2, Table 2"},{"comment":"The preprocessing is not reproducible as specified: the text says 10-core drops users with fewer than 10 interactions, but Algorithm 1 filters out items with fewer than 10 interactions; the user-retention step keeps users with the highest Jaccard distance to the full item set using a 'selected_ratio' taken from the original dataset, which substantially changes the evaluation population. The authors should release the exact processed datasets and code, and clarify whether all baseline results are produced on the same processed data; otherwise the comparison in Table 4 cannot be verified.","section":"III-A, Algorithm 1 and Section IV-A"}],"minor_comments":[{"comment":"The 10-core description is inconsistent: the text says users with fewer than 10 interactions are dropped, while Algorithm 1 filters out items with fewer than 10 interactions; these are different operations and the discrepancy should be resolved.","section":"III-A, Algorithm 1"},{"comment":"Equation (14) uses notation eS, eC, and eR that is not defined consistently with (15), and the superscripts on E^{(k-1)} versus E^{(k)} appear to be off by one relative to the propagation equations in (9).","section":"III-B-5, Eq. (14)"},{"comment":"The definitions of false positive and false negative in (16) are incorrect; false negative is an actual interaction not retrieved, not 'predicted interactions not being exist on the test set.' The ranking metrics (recall@20, NDCG@20) should be defined precisely.","section":"IV-A, Eq. (16)"},{"comment":"The text mentions 'SocialGCN' in item 3, but this model is not among the baselines and is not defined; the intended reference may be SEPT.","section":"IV-C-1"},{"comment":"The conclusion contains a placeholder citation '[?]' that should be removed or replaced.","section":"V-A"},{"comment":"There are numerous typos and inconsistencies, including 'analyst' for 'analyze' in the introduction, 'Epinion' in the Table 4 header, and 'N DGC' in equation (17).","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper appears to be a preliminary version in IEEE Access format. The central formula error in Section III-A-2 is serious but correctable, so I do not recommend rejection outright. However, the evaluation also needs to be clarified regarding train/test separation for the derived matrices, and the small reported gains require error bars or significance tests. The authors should also clarify the novelty relative to their own WiGCN model, which already uses a weighted user influence matrix. I would encourage the editor to request a corrected and rerun version with code and processed data before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper has a load-bearing bug in its central equation. Equation (6) defines C = f(WU ⊙ WI^{-1}) where WU = |Ii ∪ Ij| and WI = |Ii ∩ Ij|. Element-wise that is union/intersection, which is ≥ 1 when the sets overlap and undefined when they are disjoint. That is not the Jaccard index in [0,1] that Table 2 bins. The surrounding text also calls WI the union matrix in one place, reversing the definitions. So as written, the quantized user-correlation matrix C—the component every reported improvement is attributed to—cannot be constructed. If the intended formula was WI ⊙ WU^{-1}, that is a fixable typo, but it is the core signal of the paper.\n\nCredit where it is due: the specific combination of a quantized Jaccard user-user matrix plus social edges summed into LightGCN-style propagation is new, the modular ablation (interaction-only, social-only, full model) is informative, and the convergence-speed observation is interesting. The comparison set is broad and includes appropriate recent baselines. Those are real strengths.\n\nSoft spots beyond the equation: there is no code or data, no error bars or seeds, and the quantization thresholds in Table 2 are hand-picked. The preprocessing step that keeps users with the highest Jaccard similarity to the full item set is effectively a degree filter and is applied before the train/test split, so it can leak test-set information. The reported gains are small—recall@20 improvements of 0.5–3.6% over the best baseline—and consistent with noise. None of these are fatal on their own, but together they mean the empirical claim is not yet supported.\n\nSerious thinker? I would say no as submitted, because the internal contradiction in the central definition makes the paper incoherent on its own terms. That said, it looks like a fixable typo rather than a deep conceptual error.\n\nWho should read this: people working specifically on social GCN recommenders might find the ablation framework worth a look, but the current manuscript would need a careful rewrite before it is usable.\n\nRecommendation: I would not send this to peer review in its present form. The central signal is undefined, and the lack of code, data, and variance makes the empirical claims unverifiable. The idea is salvageable—correct the equation, release the code, add seeds—but that is major surgery, not light revision.","headline":"The central user-correlation matrix C is undefined as written (union/intersection, not Jaccard), so the paper's core signal is unreproducible; the rest is a modest, interesting ablation study that would need code, data, and variance reporting before I'd trust the gains.","tokens_in":17672,"tokens_out":4062,"would_cite":false,"duration_ms":43715,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that propagating user embeddings through interaction, Jaccard-similarity, and friendship graphs matches or beats nine baselines on recall, precision, and NDCG@20 in four datasets.","keywords":["Recommender systems","Social recommender systems","Collaborative filtering","Graph convolutional networks","Jaccard similarity","User correlation","Implicit feedback"],"falsifier":"Recompute the Table 4 numbers with the correlation matrix C removed from propagation (its normalized adjacency set to zero); if recall, precision, or NDCG@20 on any dataset stays at or above the variant that uses only interaction and social signals, then C contributes nothing and the central claim fails. A second check: the formula printed as equation (6), with WU the union size and WI the intersection size, yields values of one or larger rather than Jaccard similarities in [0,1], so the quantization in Table 2 is not directly reproducible from the text alone and requires the authors' implementation.","tokens_in":16660,"feed_emoji":"🤝","tokens_out":14476,"duration_ms":126452,"temperature":0.7,"pith_summary":"The paper proposes a graph-convolution recommender, GCCF, that learns user representations from three graphs at once: user–item interactions, a quantized Jaccard similarity graph between users based on overlapping purchase histories, and a social friendship graph. The authors claim this three-signal model reports the best or tied-best recall, precision, and NDCG@20 among nine baselines—MF, SocialMF, TrustSVD, GCMC, NGCF, WiGCN, LightGCN, SocialLGN, and SEPT—across all four datasets tested (Gowalla, Librarything, Ciao, Epinions). They also report that the correlation graph makes training converge in far fewer epochs than LightGCN-style baselines. If the paper is right, a simple sum of three normalized graph propagations, with no transformation matrices or activation functions, captures influence signals that heavier social-recommendation designs miss, and the extra user-correlation signal is cheap to compute and effective.","feed_headline":"Purchase similarity plus friendship signals top nine recommenders","feed_subtitle":"On all four test datasets, the model beats LightGCN, SEPT, and seven others while training in fewer epochs.","key_machinery":"The central mechanism is multi-graph linear embedding propagation. Every layer, the user embedding aggregates three symmetrically normalized neighborhood sums—from the interaction matrix A, the social matrix S, and the correlation matrix C—and the item embedding aggregates only from A (Equation 9); the three user signals are combined by weighted sum and, in matrix form (Equation 14), by simple addition. The new object is C: the paper computes the user-correlation signal from the Jaccard overlap of users' item sets and then applies a fixed quantization function f (Table 2) that maps overlap values into five levels, so that strong overlaps strongly influence propagation while weak overlaps are zeroed out. The final embedding is the mean over K layers, and the trainable parameters are only the initial embeddings, with the BPR loss trained by Adam.","core_discovery":"The central claim is that user embeddings should be built by linearly propagating through three symmetrically normalized adjacency matrices at the same time: the user–item interaction matrix A, the social friendship matrix S, and a user-correlation matrix C derived from the Jaccard overlap of users' item sets, with overlap values quantized into five influence levels (0, 0.005, 0.05, 0.5, 1) according to Table 2. At each layer the three user-signal embeddings are simply summed and the item embedding propagates only through A (Equations 9 and 14), and the final user and item embeddings are the mean of all layer embeddings, scored by inner product and trained with Bayesian personalized ranking. With these design choices—no feature-transformation matrices and no nonlinear activations—the model reports the best or tied-best recall, precision, and NDCG@20 among MF, SocialMF, TrustSVD, GCMC, NGCF, WiGCN, LightGCN, SocialLGN, and SEPT on Gowalla, Librarything, Ciao, and Epinions. The authors interpret this as evidence that the correlation graph contributes influence signals beyond what high-order connectivity in the interaction graph and raw friendship edges already provide, and that quantizing overlaps prevents the model from overfitting to noisy small co-occurrences.","pith_inferences":["Because C is computed once from the interaction matrix with simple matrix operations, the observed convergence speedup suggests a preprocessing-only enhancement that could be bolted onto other LightGCN-style recommenders; the paper does not test that transfer.","The quantization into five fixed levels is one regularization choice; a continuous Jaccard weight or a learned mapping from overlap to influence would separate the benefit of the similarity signal itself from the benefit of threshold smoothing.","Since the social matrix is optional, the same architecture could accept other user–user relations, such as trust scores or co-purchase at category level, in place of friendship edges; the paper leaves that substitution untested."],"forward_implications":["On Gowalla, Librarything, Ciao, and Epinions, the full model reports the best or tied-best recall, precision, and NDCG@20 among the nine baselines in Table 4; the recall gains over the best baseline are roughly 0.08 to 0.68 percentage points.","Adding the correlation matrix shortens training: on Gowalla the full model converges in 440 epochs versus 1,140 for LightGCN and 1,090 for SEPT, with similar reductions on the other datasets.","The ablations show the two extra signals play different roles: on Gowalla social friendship contributes the larger share of the improvement, while on Librarything the interaction-correlation signal matters more and social data alone slightly lowers NDCG; combining both always gives the best result.","Removing both feature-transformation matrices and nonlinear activations is again confirmed as the right simplification: SocialLGN variants that drop only one of the two perform worse, and the proposed model uses neither."],"supporting_citations":[{"why":"Establishes the NGCF hop-by-hop propagation and the layer-depth setting that the proposed model inherits and compares against.","marker":"[7]"},{"why":"Provides the LightGCN linear propagation backbone (no transformation matrices, no activations) that GCCF extends with two additional user–user graphs.","marker":"[24]"},{"why":"Introduces the weighted user-influence matrix input that the paper refines into the quantized Jaccard correlation matrix C.","marker":"[34]"},{"why":"SocialLGN is the social LightGCN baseline whose component-ablation variants justify the model's design and which GCCF must beat.","marker":"[32]"},{"why":"SEPT supplies the socially-aware self-supervised tri-training baseline and the shared-view formula the paper cites when deciding to keep S separate.","marker":"[33]"},{"why":"TrustSVD is one of the social matrix-factorization baselines in Table 4, representing trust-based SVD++ that the graph model must outperform.","marker":"[25]"},{"why":"SocialMF is the trust-propagation matrix-factorization baseline in Table 4 against which the paper measures the benefit of social signals.","marker":"[46]"}],"fun_headline_variants":["Friends and shoppers alike: graph model tops 10 recommender systems","Fusing friendship and purchase similarity for sharper recommendations","Graph recommender merges social and rating signals, beats nine baselines","New GCN combines social edges and rating correlation to improve recall"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the degree of overlap between two users' purchased or rated items carries influence information that is not already present in the purchase graph or the friendship graph, and that adding the three propagation signals by simple summation is enough to turn that influence into better recommendations.","fun_headline_variants_meta":{"raw":{"variants":["Friends and shoppers alike: graph model tops 10 recommender systems","Fusing friendship and purchase similarity for sharper recommendations","Graph recommender merges social and rating signals, beats nine baselines","New GCN combines social edges and rating correlation to improve recall"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000657,"raw_usage":{"total_tokens":3087,"prompt_tokens":1102,"completion_tokens":1985,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":718,"completion_tokens_details":{"reasoning_tokens":1914}},"tokens_in":718,"tokens_out":1985,"duration_ms":14520,"temperature":1.0,"reasoning_tokens":1914,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T11:15:04.176567+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the Table 4 numbers with the correlation matrix C removed from propagation (its normalized adjacency set to zero); if recall, precision, or NDCG@20 on any dataset stays at or above the variant that uses only interaction and social signals, then C contributes nothing and the central claim fails. A second check: the formula printed as equation (6), with WU the union size and WI the intersection size, yields values of one or larger rather than Jaccard similarities in [0,1], so the quantization in Table 2 is not directly reproducible from the text alone and requires the authors' implementation.","supporting_citations":[{"cited_title":"Neural Graph Collaborative Filtering","cited_arxiv_id":"1905.08108","evidence_quote":"Establishes the NGCF hop-by-hop propagation and the layer-depth setting that the proposed model inherits and compares against."},{"cited_title":"& Wang, M","cited_arxiv_id":null,"evidence_quote":"Provides the LightGCN linear propagation backbone (no transformation matrices, no activations) that GCCF extends with two additional user–user graphs."},{"cited_title":"& Snasel, V","cited_arxiv_id":null,"evidence_quote":"Introduces the weighted user-influence matrix input that the paper refines into the quantized Jaccard correlation matrix C."},{"cited_title":"& Zeng, J","cited_arxiv_id":null,"evidence_quote":"SocialLGN is the social LightGCN baseline whose component-ablation variants justify the model's design and which GCCF must beat."},{"cited_title":"Socially-Aware Self-Supervised Tri-Training for Recommendation","cited_arxiv_id":"2106.03569","evidence_quote":"SEPT supplies the socially-aware self-supervised tri-training baseline and the shared-view formula the paper cites when deciding to keep S separate."},{"cited_title":"& Yorke-Smith, N","cited_arxiv_id":null,"evidence_quote":"TrustSVD is one of the social matrix-factorization baselines in Table 4, representing trust-based SVD++ that the graph model must outperform."}],"review_version":1}