{"id":"64226f97-c32b-44a2-8a54-9de37fd40878","arxiv_id":"2501.16722","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A wavelet-enhanced hypergraph diffusion model with two encoders and contrastive learning reports consistent, small ranking improvements over six baselines on three recommendation datasets.","lead":"WaveHDNN is a new recommendation model that combines two kinds of hypergraph learning: one encoder that treats user-item groups as hyperedges and sends different messages to different nodes, and another that uses wavelet filters to capture local structures at multiple scales. Tests on three shopping and gaming datasets show small but consistent gains in ranking accuracy over six existing recommender baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed consistent superiority is not statistically established: no variance estimates are reported, one key table entry is internally inconsistent, and the same model's score differs between tables.","rationale":"The reader's weakest assumption identifies exactly the spot where the central claim is least secure: the fairness and integrity of the reported comparisons. My review of the full text confirms that no error bars, significance tests, code, or data splits are provided, and the tables contain demonstrable inconsistencies. This is more load-bearing than other possible concerns such as the absence of ED-HNN and wavelet HGCN as baselines, because the central claim as literally stated ('outperforms all baselines') concerns the listed baselines only; the missing source-model comparisons weaken significance but do not directly falsify the claim. The internal errors and lack of statistical support do directly threaten the claim, and they are easy to verify and fix. The verdict should remain CONDITIONAL: the empirical claims need to be backed by corrected tables, per-run statistics, and released artifacts before the paper's central assertion can be accepted. I agree with the reader's assessment.","tokens_in":13024,"tokens_out":5751,"duration_ms":59179,"concrete_test":"Re-run the full comparison with the same 7:1:2 splits and at least 10 random seeds; report per-metric mean, standard deviation, and a paired Wilcoxon signed-rank test for WaveHDNN vs each baseline. Also verify the SHT Steam NDCG@20 entry and the Amazon Recall@40/NDCG@40 values in Tables 2 and 5. If the corrected SHT value or the seed variance changes any ranking, or if WaveHDNN's Steam NDCG@10 margin is not significant at p<0.05, the 'consistently outperforms' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that WaveHDNN 'consistently outperforms all baselines across six evaluation metrics' on three datasets. For that claim to hold, the ordinal rankings in Tables 2-4 must be reliable. Three problems jointly undermine this. First, no standard deviations or significance tests are given; with only 5 averaged runs, a relative gain as small as 1.59% (Steam NDCG@10, WaveHDNN 0.07754 vs AutoCF 0.07632) can easily be within run-to-run noise. Second, Table 3 reports SHT NDCG@20 = 0.02647, far below its NDCG@10 = 0.07583 and inconsistent with its Recall@20 = 0.12687; this indicates a data error in a table that the superiority argument depends on. Third, WaveHDNN's Amazon Recall@40 is 0.21812 in Table 2 but 0.21845 in Table 5 (and NDCG@40 differs as well), so even the paper's own reported numbers are not internally consistent. Without code, data splits, or per-baseline tuning budgets, a reader cannot verify whether the reported margins are real or artifacts of favorable settings. The ablation does show both components contribute, but it does not repair the central empirical comparison. These reproducibility and reporting defects are load-bearing because the paper's primary contribution is an empirical state-of-the-art claim, not a theoretical guarantee.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes WaveHDNN, a hypergraph-based collaborative filtering model with two parallel encoders: a Heterophily-aware Collaborative Encoder inspired by ED-HNN, and a Multi-scale Group-wise Structure Encoder using wavelet hypergraph convolution. The two views are aligned with an InfoNCE-style cross-view contrastive loss, and the final score is computed from the concatenated embeddings. The authors report experiments on Amazon-Books, Steam, and Yelp, comparing against LightGCN, SGL, DHCF, HCCF, SHT, and AutoCF across Recall@10/20/40 and NDCG@10/20/40, and they report ablations removing each encoder.","tokens_in":13270,"tokens_out":6501,"duration_ms":58124,"significance":"If the empirical claims hold, WaveHDNN would be a state-of-the-art hypergraph collaborative filtering method, with relative gains over the second-best baseline of roughly 3–7% on the three datasets. The architectural combination of heterophily-aware message passing and multi-scale wavelet filtering is well motivated, and the ablations in Table 5 provide initial evidence that both components contribute. The paper also transparently builds on external works (ED-HNN and wavelet HGCN). However, the strength of the paper currently rests entirely on the reliability of the reported numbers, which are not statistically supported and contain internal inconsistencies. The paper ships no code, no per-baseline tuning budgets, and no variance estimates, all of which are needed to verify the central empirical superiority claim.","major_comments":[{"comment":"The SHT row reports NDCG@20 = 0.02647, which is about one-third of its NDCG@10 = 0.07583, while Recall@20 is 0.12687. With non-negative relevance gains, NDCG@20 cannot drop that far below NDCG@10; this entry is internally inconsistent and suggests a data transcription error. Because Tables 2–4 carry the central claim of consistent superiority, this error must be corrected and the affected conclusions re-verified.","section":"Table 3 (Steam)"},{"comment":"The WaveHDNN row gives Amazon Recall@40 = 0.21812 in Table 2 but 0.21845 in Table 5, and NDCG@40 = 0.12484 in Table 2 but 0.12486 in Table 5, for what is described as the same model and the same 5-run average. The discrepancy implies that at least one of these tables does not reflect the final configuration, and it undermines confidence in the reproducibility of the headline numbers.","section":"Tables 2 and 5"},{"comment":"The paper reports only the mean over 5 runs with no standard deviations, confidence intervals, or significance tests. The smallest reported margin in the main tables is 1.59% (Steam NDCG@10: WaveHDNN 0.07754 vs. AutoCF 0.07632), which is likely within run-to-run variation; the claim that WaveHDNN \"consistently outperforms all baselines across six evaluation metrics\" is therefore not statistically established. Please provide per-metric variance and paired significance tests over the 5 runs.","section":"§4.1 and Tables 2–4"},{"comment":"The experimental protocol is under-specified: the paper does not state the embedding dimension, learning rate, batch size, regularization, number of layers for each model, the temperature τ in Eq. (4), or the per-baseline hyperparameter tuning budget. Without these details and without code or data splits, a reader cannot assess whether the reported margins reflect a fair comparison or favorable configuration. This is a load-bearing gap for an empirical state-of-the-art claim.","section":"§4.1–4.2"},{"comment":"The wavelet-based hypergraph convolution is not fully defined: the wavelet basis Θ, its inverse Θ′, and the filter matrix Λ are introduced without specifying how the hypergraph wavelet transform is computed or how Λ is parametrized and learned. The reference to [14] is not sufficient because the paper modifies the layer with a residual concatenation, and the reader cannot implement or reproduce the proposed encoder from the text. Please supply the complete formulation or an unambiguous pointer to the exact construction used.","section":"§3.2, Eq. (3)"}],"minor_comments":[{"comment":"There are typos: \"datsets\" should be \"datasets\", and the section opening \"mong\" should be \"Among\".","section":"§4.5"},{"comment":"The model name appears as \"W aveHDNN\" in several table rows; please remove the spurious space.","section":"Tables 2–4"},{"comment":"The phrase \"The top mark performance is highlighted\" is awkward; consider \"The best performance is in bold\" and similar idiomatic wording.","section":"Table captions"},{"comment":"AutoCF is listed as a baseline but no citation number is provided in the text; please add the corresponding reference.","section":"§4.2"},{"comment":"The summation bounds IX and LX in Eq. (4) are not defined, and the layer index l is ambiguous; please specify the ranges and clarify how layer-wise contrastive losses are aggregated.","section":"Eq. (4)"},{"comment":"The line \"Preprint submitted to Technical Report\" is not appropriate for a journal submission and should be removed.","section":"Header"}],"recommendation":"major_revision","confidential_remarks":"The manuscript has the flavor of an extended technical report and will need to meet normal journal reporting standards. The empirical verification issues are addressable in principle, but they are critical: without corrected tables, variance estimates, and a fully specified experimental protocol, the central superiority claim is not defensible. If the authors cannot resolve the internal inconsistencies or add the missing statistical support, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"WaveHDNN is a reasonable engineering contribution: it takes the equivariant operator from ED-HNN, the wavelet hypergraph convolution from [14], and cross-view InfoNCE, and applies the combination to collaborative filtering. That is a real combination, not a trivial one, and the ablation in Table 5 does what an ablation should: removing either encoder clearly hurts. The paper is also honest about its building blocks, which I appreciate.\n\nBut the central claim, that WaveHDNN “consistently outperforms all baselines,” isn’t yet supported cleanly. Table 3 has at least one impossible entry: SHT’s NDCG@20 is 0.02647, far below its own NDCG@10 of 0.07583, and AutoCF’s NDCG@10 (0.07632) is higher than its NDCG@20 (0.07615). Those look like transposition or rounding errors, but they make the whole Steam table hard to trust. Table 5 also gives WaveHDNN’s Amazon Recall@40 as 0.21845 while Table 2 says 0.21812; small, but it shows the numbers weren’t double-checked. And there are no standard deviations or significance tests anywhere. Some of the reported margins, like the 1.59% gain on Steam NDCG@10, are exactly the size that run-to-run noise can explain. With no code, no data splits, and no per-baseline tuning budget, a reader can’t reproduce the setup.\n\nThe stress-test note is therefore on target, and I agree with the conditional verdict. Still, I’d hesitate to call these load-bearing flaws in the sense of a wrong method. The architecture is sensible, the ablation supports both components, and the direction of the gains is consistent across datasets. The problems are reporting and reproducibility problems, not evidence of a broken approach. What’s missing is a cleaner, statistically grounded comparison and artifacts. I also would have liked to see ED-HNN and the wavelet HGCN themselves as baselines, since the novelty claim rests on beating the source models, not just LightGCN and AutoCF.\n\nThe paper deserves a serious referee. It is not a desk reject; it is a revise-and-resubmit case. I’d send it out, require corrected tables, standard deviations, and released code or splits, and ask for a comparison against the two cited source components. If the numbers hold up after that, it’s a modest but useful result for hypergraph-based CF.","headline":"A plausible, incrementally novel hypergraph CF model with a sensible two-encoder design, but the empirical superiority claim is undercut by inconsistent tables and missing uncertainty estimates.","tokens_in":13845,"tokens_out":1841,"would_cite":false,"duration_ms":19369,"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":"WaveHDNN, a wavelet-hypergraph diffusion recommender, claims consistent wins over six baselines on three datasets across all reported metrics.","keywords":["recommender systems","collaborative filtering","hypergraph diffusion","heterophily","wavelet transform","multi-scale structure encoding","contrastive learning","over-smoothing"],"falsifier":"Re-run WaveHDNN and the six baselines on the same three datasets with an equal hyperparameter budget and report per-seed standard deviations; if any baseline ties or beats WaveHDNN on any of the six metrics, the uniform superiority claim fails. A cheaper check is to reproduce Table 3's SHT NDCG@20 entry, because the printed value (0.02647) sits below that model's NDCG@10 value (0.07583), and to check Table 5's WaveHDNN Amazon Recall@40 (0.21845), which differs from Table 2's value (0.21812).","tokens_in":12760,"feed_emoji":"📈","tokens_out":8401,"duration_ms":71246,"temperature":0.7,"pith_summary":"This paper introduces WaveHDNN, a collaborative-filtering model that treats user-item interactions as a hypergraph and propagates information through wavelet-based diffusion. It claims that WaveHDNN outperforms six existing graph- and hypergraph-based recommender baselines on Amazon-Books, Steam, and Yelp across all six reported metrics, Recall@10/20/40 and NDCG@10/20/40. The design separates two failure modes the authors say limit prior GNN recommenders: heterophilic interactions, where a user's items span many categories, and over-smoothing, where deep layers make embeddings indistinguishable. A heterophily-aware encoder handles the first, a multi-scale wavelet structure encoder handles the second, and cross-view contrastive learning aligns them. If the comparisons are fair, the result is a practical recipe for higher-order group-wise recommendation with consistent gains over strong baselines.","feed_headline":"Hypergraph diffusion model WaveHDNN tops baselines on all six metrics","feed_subtitle":"Two encoders, one for heterophily and one for wavelet multi-scale structure, lift Recall and NDCG by up to 7.24% on three benchmarks.","key_machinery":"The load-bearing mechanism is a two-channel hypergraph diffusion architecture. Channel one, the Heterophily-aware Collaborative Encoder, transforms node embeddings with MLPs, passes them through a hypergraph convolution layer, applies layer normalization, and adds the result back to the original embeddings; this equivariant residual design lets messages to nodes sharing a hyperedge differ rather than converge. Channel two, the Multi-scale Group-wise Structure Encoder, is a wavelet hypergraph convolution $X^{(l+1)} = \\Theta \\Lambda \\Theta' X^{(l)} W + X^{(l)}$, where $\\Theta$ and $\\Theta'$ are the wavelet basis and its inverse and $\\Lambda$ is a diagonal frequency filter; this is what localizes message passing to specific regions and scales of the hypergraph. The two channels are tied together by an InfoNCE-style cross-view contrastive loss and trained with a BPR ranking loss.","core_discovery":"The paper's central claim is that a single model can handle heterophilic user-item interactions and localized high-order structure at the same time, and that doing so yields consistently better recommendations than existing graph- and hypergraph-based collaborative filtering. WaveHDNN encodes users and items through two separate channels. The Heterophily-aware Collaborative Encoder uses an equivariant operator, built on hypergraph diffusion, to send different messages to different nodes within a hyperedge, then layer-normalizes and adds the input back to preserve identity. The Multi-scale Group-wise Structure Encoder applies wavelet hypergraph convolution, $X^{(l+1)} = \\Theta \\Lambda \\Theta' X^{(l)} W + X^{(l)}$, so the spread of information can be tuned per scale instead of by stacking more layers. A cross-view contrastive loss keeps the two encoders' embeddings consistent, and ranking is optimized with a Bayesian personalized ranking loss. On the three datasets the model reports wins on all six metrics for every baseline, with the largest relative margin, 7.24%, on Steam NDCG@20; ablations show removing either encoder lowers Recall@40 and NDCG@40 on all three datasets.","pith_inferences":["An independent re-run with matched hyperparameter budgets and per-seed variances would be the decisive test; the paper reports point estimates only, so the ordinal claim is exactly what needs reproduction.","The wavelet scale parameters could be adapted to temporal recommender settings, where local neighborhoods drift over time, but the paper only evaluates static datasets.","The equivariant heterophily encoder is not limited to recommendation; the same message-differentiation mechanism could apply to hypergraph node classification or link prediction in graphs with mixed homophily."],"forward_implications":["Hypergraph collaborative filtering can capture both heterophily and multi-scale locality without very deep GNN stacks, because the wavelet channel tunes information spread by scale instead of by layer count.","On the three datasets tested, the reported gains are consistent across all six ranking metrics, not concentrated in one cutoff, suggesting the design improves overall ranking quality rather than only recall.","Ablations say both channels matter: removing either encoder drops Recall@40 and NDCG@40 on Amazon-Books, Steam, and Yelp, so the improvement is not carried by a single component.","The largest relative margin appears on Steam at 7.24% NDCG@20, so the method may be most valuable in denser interaction data where group-wise structure is more informative."],"supporting_citations":[{"why":"Supplies the equivariant hypergraph diffusion operator that the Heterophily-aware Collaborative Encoder builds on.","marker":"[13]"},{"why":"Supplies the wavelet-based hypergraph transform used in the Multi-scale Group-wise Structure Encoder.","marker":"[14]"},{"why":"HCCF is a hypergraph contrastive collaborative filtering baseline and a key comparison for the proposed model.","marker":"[7]"},{"why":"SHT is the self-supervised hypergraph transformer baseline that serves as the strongest hypergraph competitor in the tables.","marker":"[53]"},{"why":"LightGCN is the simplified graph convolution baseline that WaveHDNN must beat.","marker":"[5]"},{"why":"DHCF is a dual-channel hypergraph collaborative filtering baseline used for comparison.","marker":"[52]"},{"why":"Supplies the InfoNCE-style contrastive objective used for cross-view embedding consistency.","marker":"[64]"},{"why":"Supplies the BPR ranking loss used to optimize WaveHDNN's recommendation score.","marker":"[65]"}],"fun_headline_variants":["WaveHDNN: hypergraph diffusion beats GNNs on all metrics","Wavelet hypergraph diffusion lifts Recall and NDCG by 7.24%","WaveHDNN harnesses heterophily and wavelets for top recsys","Hypergraph diffusion with wavelets wins six recommendation metrics","WaveHDNN: one model, two encoders, top Recall and NDCG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that WaveHDNN consistently outperforms every baseline rests on the fairness and accuracy of the reported comparisons, which the paper supports only with point estimates and no tuning budget or code.","fun_headline_variants_meta":{"raw":{"variants":["WaveHDNN: hypergraph diffusion beats GNNs on all metrics","Wavelet hypergraph diffusion lifts Recall and NDCG by 7.24%","WaveHDNN harnesses heterophily and wavelets for top recsys","Hypergraph diffusion with wavelets wins six recommendation metrics","WaveHDNN: one model, two encoders, top Recall and NDCG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001142,"raw_usage":{"total_tokens":4776,"prompt_tokens":1016,"completion_tokens":3760,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":632,"completion_tokens_details":{"reasoning_tokens":3661}},"tokens_in":632,"tokens_out":3760,"duration_ms":24426,"temperature":1.0,"reasoning_tokens":3661,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T11:10:26.950556+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run WaveHDNN and the six baselines on the same three datasets with an equal hyperparameter budget and report per-seed standard deviations; if any baseline ties or beats WaveHDNN on any of the six metrics, the uniform superiority claim fails. A cheaper check is to reproduce Table 3's SHT NDCG@20 entry, because the printed value (0.02647) sits below that model's NDCG@10 value (0.07583), and to check Table 5's WaveHDNN Amazon Recall@40 (0.21845), which differs from Table 2's value (0.21812).","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet-based hypergraph transform used in the Multi-scale Group-wise Structure Encoder."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"HCCF is a hypergraph contrastive collaborative filtering baseline and a key comparison for the proposed model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SHT is the self-supervised hypergraph transformer baseline that serves as the strongest hypergraph competitor in the tables."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"LightGCN is the simplified graph convolution baseline that WaveHDNN must beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DHCF is a dual-channel hypergraph collaborative filtering baseline used for comparison."},{"cited_title":"Rendle, C","cited_arxiv_id":null,"evidence_quote":"Supplies the BPR ranking loss used to optimize WaveHDNN's recommendation score."}],"review_version":1}