{"id":"fd067227-460d-443d-a343-323331ee11b5","arxiv_id":"2505.24172","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MCL improves heterogeneous-graph recommendation accuracy and robustness by combining random masking and random propagation with cross-view contrastive learning on one-hop and meta-path neighborhoods.","lead":"Recommendation models that use heterogeneous graphs often amplify noise along meta-paths. This paper adds random masking and contrastive learning between local and high-order views, and reports consistent accuracy gains on three standard datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The contrastive positive-matrix definition in §III-C2 is internally ambiguous and, on dense meta-paths such as UAU/UMGMU, may either manufacture false positives or secretly rely on oracle meta-path selection, which would undermine the Fig. 4 robustness claim.","rationale":"The reader's weakest assumption is the same one I identify: meta-path connectivity is used as a positive-pair proxy without validating its reliability, and dense or noisy meta-paths make this proxy dangerous. The paper itself contains strong evidence for this concern: in §IV-C2 it states that UAU has only nine nodes and that 'many users inevitably interact through UAU', calling this a source of noise. That admission makes the §III-C2 definition of positives 'based on all meta-paths' decisive either way. If all meta-paths are truly used, the contrastive loss will reinforce the noisy UAU/UMGMU connections and cannot be the denoising mechanism claimed. If only original meta-paths are used, then MCL's robustness in Fig. 4 is partially an oracle result, because the model is told which meta-paths are useful while the baselines are not. This is a genuine fork in the argument, not a mere reproducibility gap. I do not see a fatal flaw in the clean-data ranking experiments: MCL's Table II gains and ablations are plausible, and the method may still work through the masking/propagation component. But the robustness claim is the paper's headline contribution, and it rests on unresolved ambiguity plus missing code and single-seed results, so the reader's CONDITIONAL verdict is appropriate. I therefore leave the verdict unchanged while insisting on the positive-matrix density check as the key to deciding whether MCL's robustness conclusion survives.","tokens_in":18577,"tokens_out":6849,"duration_ms":90946,"concrete_test":"Run the §IV-C2 redundant-meta-path experiment with two MCL variants: (A) Pos_u/Pos_i built from all available meta-paths (including UAU, UOU, UMGMU) and (B) Pos_u/Pos_i built only from the original meta-path set. Report the off-diagonal density of Pos_u/Pos_i in each variant and recall@20/NDCG@20 on all three datasets. If variant A's positive-matrix density exceeds 0.5 on MovieLens (where UAU alone connects most users) and variant B reproduces Fig. 4's small declines while A does not, then the robustness result is driven by oracle selection of positive meta-paths rather than by random masking or contrastive denoising.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-C2 defines Pos_u/Pos_i by assigning 1 to pairs connected by 'all meta-paths', yet the same section's example excludes UMGMU-connected users from positive pairs. Section IV-C2 admits that UAU has only nine intermediate nodes, so 'many users inevitably interact through UAU', and describes such fused meta-paths as noise. If UAU/UMGMU enter Pos, the contrastive loss aligns the very false-positive pairs the model claims to be robust against; if they do not enter Pos, MCL silently uses only the hand-picked, validation-selected original meta-paths to build positives, so Fig. 4 compares MCL (given the clean meta-path set) against baselines that must use all meta-paths. The noisy-interaction experiment (Fig. 3) is similarly affected: randomly added U-I edges alter UMU connectivity and therefore Pos_u. The manuscript reports neither the density of Pos_u/Pos_i nor the exact meta-path set used to build them, leaving the central robustness mechanism untestable as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes Masked Contrastive Learning (MCL), a heterogeneous graph recommender that augments user and item embeddings by randomly masking entire embedding vectors with Bernoulli masks and averaging L random-propagation steps over meta-path-based homogeneous subgraphs. It then builds two views, one from one-hop relational attention and one from meta-path node/semantic attention, and aligns them with a contrastive loss whose positive pairs are derived from a meta-path-based user-user and item-item interaction matrix. The final objective combines the two contrastive terms with a BPR loss. Experiments on MovieLens-100k, Amazon, and Yelp compare MCL with HAN, HGT, HeCo, RoHe, HERec, KGAT, NGCF+, SMIN, and HGCL, and include noise-injection and redundant-meta-path robustness tests.","tokens_in":18824,"tokens_out":5279,"duration_ms":62028,"significance":"If the robustness claims hold, MCL offers a simple and transferable recipe: random mask plus random propagation as a graph-agnostic augmentation, and cross-view contrastive learning between one-hop and meta-path views. The paper's breadth of baselines and three datasets is a strength, and the ablation study isolates the mask, contrastive, one-hop, and meta-path components. The central empirical claim, however, depends on the exact construction of the positive-pair matrix and on the configuration of the robustness experiments, both of which are underspecified. The value of the contribution can be assessed only after these issues are resolved with concrete reporting.","major_comments":[{"comment":"The definition of Pos_u and Pos_i is internally inconsistent. The text states that positive samples receive a value of 1 'based on all meta-paths that interact with each other', but the immediately following example excludes a pair connected by 'U-M-G-M-U' because the pair is not connected by 'U-M-U'. It must be specified whether a positive pair requires connectivity under every meta-path in the chosen set, under at least one meta-path, or under a manually selected subset. This choice determines whether dense meta-paths such as UAU and UMGMU enter the contrastive objective, so the robustness conclusions in Section IV-C2 cannot be evaluated without resolving this ambiguity.","section":"III-C2, Eqs. (8)-(9)"},{"comment":"The redundant-meta-path experiment does not state whether the added meta-paths from Table IV are included in Pos_u and Pos_i for MCL. If they are included, the contrastive loss aligns pairs created by the very dense paths that the paper itself describes as noise; if they are excluded, MCL is tested with clean, validation-selected meta-paths while the baselines must use all meta-paths. The authors should report the exact meta-path set used to build the positive matrices, the density of Pos_u and Pos_i under the original and augmented settings, and how the noisy-interaction experiment in Fig. 3 changes Pos_u when random user-item edges are injected.","section":"IV-C2, Fig. 4 and Table IV"},{"comment":"No error bars, multiple seeds, or significance tests are reported in Tables II, III, and V or in Figs. 3-4, although several reported improvements are only 1-5% relative and could be within seed variance. In addition, the final values of the mask retain probability p, the propagation order L, the contrastive balance beta, the view balance lambda1, the temperature tau, and the learning rate are never listed; Section IV-E describes tuning ranges but not the selected settings. The experiments are therefore not reproducible as written.","section":"IV-A.3 and IV-E"},{"comment":"The contrastive objective is not well-defined as presented. Eq. (7) applies softmax to a scalar expression exp(sim(...)/tau), score_u is defined in Eq. (8) as Pos_u * sim_u, but score^1_u and score^2_u in Eq. (9) are never defined, and the relation between sim^1_u, sim^2_u, Pos_u, and the two views is unclear. Please provide exact matrix dimensions and coordinate-wise definitions for all quantities in the loss.","section":"III-C1, Eqs. (6)-(9)"}],"minor_comments":[{"comment":"In Eq. (2), the normalized adjacency is written as (D^{-1/2})^2, which is dimensionally impossible; it should presumably be D^{-1/2} A D^{-1/2}.","section":"Eq. (2)"},{"comment":"References [54]-[59] appear in the bibliography but are never cited in the body; several are same-author papers closely related to the proposed method, so they should be either cited where relevant or removed.","section":"References"},{"comment":"The model is introduced as 'Masked Contrastive Learning' throughout the paper, but the conclusion refers to 'Mask Contrastive Learning'; please make the name consistent.","section":"Conclusion"},{"comment":"The first sentence of Section III-A contains the typo 'offen' for 'often'.","section":"III-A"},{"comment":"The caption of Fig. 3 says the histogram represents recall@20 and the line graphs represent NDCG@20, but the figure appears to use bars and line markers interchangeably; please align the caption with the figure.","section":"Fig. 3"}],"recommendation":"major_revision","confidential_remarks":"The uncited references [54]-[59] are all from the same research group and overlap with the paper's topic; I would ask the authors to cite or remove them and to disclose prior work clearly. No code is released, and the missing hyperparameter values and positive-matrix ambiguity make reproduction difficult. I would not recommend acceptance before a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. MCL is a coherent, incremental combination of known ideas — random node masking, feature propagation without learnable weights, two-view contrastive learning, and meta-path-aware positive sampling — and its reported gains over strong baselines like HGCL and HeCo are real-looking but modest (roughly 1-10% relative). The robustness story is the most interesting part, but the paper as written is not fully reproducible, and the positive-matrix definition is ambiguous enough that one of the two readings would undercut the headline robustness claim.\n\nWhat is genuinely new: the specific recipe itself — whole-embedding masking before random propagation on meta-path subgraphs, contrasted against the one-hop view with a positive matrix built from meta-path connectivity — does not appear in HeCo, SGL, SimGCL, or HGCL. The experiments are broad: three standard datasets, ten baselines, separate robustness tests for noisy interactions and redundant meta-paths, plus ablations showing each component earns its keep. The mask component in particular gives the largest gain on two of three datasets, which is a real signal that the perturbation trick matters.\n\nWhere the soft spots are. First, Section III-C2's Pos matrix. The text says positives are assigned 'based on all meta-paths that interact with each other,' then gives an example where users connected via U-M-G-M-U but not U-M-U are not positive. That implies intersection across the meta-path set, not union. Under intersection, the stress-test worry about dense UAU/UMGMU manufacturing false positives doesn't land — adding a dense meta-path to an intersection prunes positives rather than adding them. But the paper never states this explicitly, never reports the density of Pos_u/Pos_i, and never says which meta-path set is used in the original versus robustness runs. If the intersection is near-empty, the contrastive loss has little signal; if the implementation actually uses union, then Fig. 4 is not testing what the text claims. A referee will need this pinned down.\n\nSecond, reproducibility. No code, no multiple seeds, no error bars, and the final values for mask probability p, propagation order L, beta, lambda1, tau, and learning rate are not reported. The hyperparameter section talks about ranges but not chosen values, so Table II is not repeatable as written.\n\nThird, references [54]-[59] are same-author papers never cited in the body, and at least one is extremely close in topic. That is the most serious concern because it affects novelty and attribution; the authors need to cite and differentiate them.\n\nThe claims are empirical, not derived from a hidden identity, so there is no circularity problem beyond the usual tuning risk. This paper is for people working on heterogeneous graph recommendation, contrastive learning, or robustness in recommenders. They would get a useful recipe and a plausible baseline. It is not a field-changer.\n\nRecommendation: accept for peer review. The architecture is coherent, the experiments are broad, and the robustness angle is worth examining. But my verdict is conditional — I would want the missing details, the Pos clarification, and the self-reference cleanup before trusting the numbers.","headline":"A plausible but underspecified incremental HGNN+CL recipe with interesting robustness results; reproducibility and uncited same-author references need fixing before trusting the numbers.","tokens_in":19336,"tokens_out":8723,"would_cite":false,"duration_ms":96254,"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":"Random masking of meta-path neighborhoods, paired with cross-view contrastive learning, is claimed to make heterogeneous graph recommenders accurate even as noisy and redundant edges are added.","keywords":["heterogeneous information network","recommendation","meta-path","contrastive learning","random masking","graph neural network","robustness","noise"],"falsifier":"Construct a deliberately meaningless meta-path by connecting users through a randomly permuted attribute that carries no preference signal, then run MCL with the same hyperparameters; if recall@20 and NDCG@20 drop as much as the unregularized HAN or RoHe baselines when that meta-path is added, the claimed robustness to redundant meta-paths is not real. A complementary check is to replace the multi-path positive matrix with random user-user positives: if performance does not collapse, the sampling mechanism, not the mask, is doing the work.","tokens_in":1704,"feed_emoji":"🛡️","tokens_out":2057,"duration_ms":82719,"temperature":0.7,"pith_summary":"Heterogeneous graph recommenders enrich sparse user-item data with meta-path connections, but those extra edges are often dense and noisy, and graph propagation magnifies the noise across many nodes. This paper proposes Masked Contrastive Learning (MCL), which randomly masks whole user and item embeddings, propagates them over meta-path subgraphs, and then contrasts the resulting meta-path view against a one-hop neighbor view. The paper argues that this makes learned embeddings less dependent on any single neighbor, so recommendation quality declines only slightly when noisy interactions or redundant meta-paths are introduced. On MovieLens, Amazon, and Yelp, MCL reports consistent gains over prior heterogeneous graph and contrastive recommenders in Recall@K and NDCG@K. If the paper is right, adding extra meta-paths no longer forces a large performance penalty, which would remove a major practical obstacle to using heterogeneous graphs in recommender systems.","feed_headline":"Mask meta-path noise and keep recommendation accuracy","feed_subtitle":"Contrastive learning over masked subgraphs beats prior graph recommenders on three real-world datasets.","key_machinery":"The load-bearing mechanism is the mask-then-propagate augmentation of user and item embeddings, combined with cross-view contrastive alignment. A Bernoulli mask multiplies each node's embedding by a 0/1 draw; the masked embeddings are then passed through L rounds of degree-normalized random propagation on the meta-path subgraphs, $\\bar A = D^{-1/2} A D^{-1/2}$, and averaged across rounds. The resulting augmented embedding feeds into node-level and semantic-level attention, while a one-hop relational-attention branch produces a local view. The two projected views are contrasted through a similarity matrix in which positive entries are user-user or item-item pairs reachable by any meta-path, scored by $\\mathrm{softmax}(\\exp(\\mathrm{sim}/\\tau))$ and optimized by a log-loss that balances the two views. This mechanism is what lets the model keep high-order structure while dampening sensitivity to individual noisy edges.","core_discovery":"MCL's central claim is that a noisy meta-path subgraph can be made usable for recommendation by perturbing the input embeddings rather than by cleaning the graph. User and item embeddings are first masked with Bernoulli draws, then randomly propagated over the meta-path subgraphs for L steps and averaged, so no single neighbor is critical to a node's representation. A two-level attention network aggregates these augmented embeddings per meta-path, while a parallel relational-attention branch aggregates one-hop heterogeneous neighbors. Contrastive learning then aligns the two views, using a multi-path sampling matrix that declares any user-user or item-item pair connected by at least one meta-path as a positive pair and all others as negatives, with a shared projection head and a temperature-scaled cosine similarity. The paper reports that this design outperforms prior heterogeneous graph recommendation and contrastive learning baselines on three datasets and, crucially, that its performance declines only slightly when 10% or 30% noise edges are injected or when additional redundant meta-paths are added.","pith_inferences":["A natural extension the paper does not test is to weight meta-path positives by estimated reliability rather than treating every meta-path connection as a binary positive, which could help when one meta-path is much noisier than others.","The same mask-and-propagate augmentation could serve as a general denoising operator for other heterogeneous graph tasks, such as node classification, where dense relation subgraphs are also common.","The paper's robustness tests cover injected interaction noise and redundant meta-paths but not adversarial edge attacks; a targeted-perturbation experiment would delimit how far the noise-resistance claim extends.","The reported embedding visualization suggests a balance between clustering and uniformity that may reduce popularity bias, but the paper does not measure popularity bias directly; an evaluation on long-tail items would make that consequence explicit."],"forward_implications":["If MCL's results hold, the most effective way to handle dense meta-path subgraphs may be to inject Bernoulli masking and random propagation before attention aggregation rather than to prune the graph explicitly.","The cross-view contrastive loss with meta-path-based positives yields a stronger self-supervised signal than generic InfoNCE applied to user-item interactions, as the paper's comparison against an InfoNCE variant indicates.","Sparse recommendation scenarios, where added noise harms graph structure most, stand to benefit most from this robustness behavior, with the largest reported relative gain on the sparsest dataset, Yelp.","The mask component is the largest single contributor to the reported gains; removing it in the ablation hurts performance more than removing contrastive learning or either of the two views.","Manual meta-path selection becomes less critical, because adding all reasonable meta-paths no longer forces a large performance penalty for MCL."],"supporting_citations":[{"why":"Supplies the cross-view co-contrastive mechanism that MCL adapts to recommendation, and serves as a baseline.","marker":"[14]"},{"why":"Gives the heterogeneous graph contrastive recommendation baseline and the InfoNCE comparison used in Table III.","marker":"[18]"},{"why":"Supplies the self-supervised graph contrastive learning recipe of random augmentations and InfoNCE that MCL extends with meta-path positives.","marker":"[19]"},{"why":"Defines the two-level meta-path attention backbone and is a principal baseline in the experiments.","marker":"[13]"},{"why":"Provides the robustness-oriented heterogeneous GNN baseline and motivates the perturbation-enlargement problem.","marker":"[15]"},{"why":"Defines meta-paths and meta-path-based similarity, the structural basis of the positive sampling matrix.","marker":"[25]"},{"why":"Provides the Bayesian Personalized Ranking pairwise loss used as the final recommendation objective.","marker":"[28]"}],"fun_headline_variants":["Mask embeddings, not edges, for robust recommendation","Random masking of embeddings tames noisy meta-paths","Contrastive views on masked meta-path graphs beat cleaning","Noisy meta-paths? Mask embeddings and contrastively learn"],"cache_read_input_tokens":21504,"weakest_assumption_plain":"The result collapses if meta-path connectivity is not a reliable proxy for user-user or item-item similarity, because the contrastive loss would then treat false pairs as positives and train the model to align unrelated nodes.","fun_headline_variants_meta":{"raw":{"variants":["Mask embeddings, not edges, for robust recommendation","Random masking of embeddings tames noisy meta-paths","Contrastive views on masked meta-path graphs beat cleaning","Noisy meta-paths? Mask embeddings and contrastively learn"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000497,"raw_usage":{"total_tokens":2415,"prompt_tokens":904,"completion_tokens":1511,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":1455}},"tokens_in":520,"tokens_out":1511,"duration_ms":11922,"temperature":1.0,"reasoning_tokens":1455,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:31:47.744326+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a deliberately meaningless meta-path by connecting users through a randomly permuted attribute that carries no preference signal, then run MCL with the same hyperparameters; if recall@20 and NDCG@20 drop as much as the unregularized HAN or RoHe baselines when that meta-path is added, the claimed robustness to redundant meta-paths is not real. A complementary check is to replace the multi-path positive matrix with random user-user positives: if performance does not collapse, the sampling mechanism, not the mask, is doing the work.","supporting_citations":[{"cited_title":"Self-supervised heterogeneous graph neural network with co-contrastive learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the cross-view co-contrastive mechanism that MCL adapts to recommendation, and serves as a baseline."},{"cited_title":"Heterogeneous graph contrastive learning for recommendation,","cited_arxiv_id":null,"evidence_quote":"Gives the heterogeneous graph contrastive recommendation baseline and the InfoNCE comparison used in Table III."},{"cited_title":"Self- supervised graph learning for recommendation,","cited_arxiv_id":null,"evidence_quote":"Supplies the self-supervised graph contrastive learning recipe of random augmentations and InfoNCE that MCL extends with meta-path positives."},{"cited_title":"Robust heterogeneous graph neural networks against adversarial attacks,","cited_arxiv_id":null,"evidence_quote":"Provides the robustness-oriented heterogeneous GNN baseline and motivates the perturbation-enlargement problem."},{"cited_title":"Pathsim: Meta path-based top-k similarity search in heterogeneous information networks,","cited_arxiv_id":null,"evidence_quote":"Defines meta-paths and meta-path-based similarity, the structural basis of the positive sampling matrix."}],"review_version":1}