{"id":"cbc8079e-146e-4ed4-87c2-13c873d56f11","arxiv_id":"2504.18588","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A snowflake-core tensor factorization model achieves slightly lower prediction error than three baselines for dynamic QoS data, but the reported gains are tiny and the experiments lack statistical validation.","lead":"This paper proposes a non-negative tensor factorization model with a snowflake-shaped core tensor to predict missing quality-of-service (QoS) values across users, services, and time. It reports small accuracy gains over three tensor completion baselines on the WS-DREAM datasets, but provides no code, hyperparameters, or error bars.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Empirical superiority rests on tiny margins with no error bars, repeated runs, or hyperparameter disclosure; on D1/D2 the gains over M3 are 0.1–0.3%, so run-to-run noise could reverse the ranking.","rationale":"I read the paper as an empirical claim: NSFT yields lower MAE/RMSE than three tensor-completion baselines on four WS-DREAM splits. The strongest version of that claim requires the reported margins to be reliable, not merely positive. The reader's weakest assumption focuses on whether Eq. (12) actually minimizes Eq. (9). That is a real problem: the equations are garbled, the learning-rate cancellation is heuristic, and no convergence proof is offered, so the results are not independently verifiable from the text. However, even granting the optimization, the numerical evidence is fragile. The D1/D2 improvements over the closest baseline are around 0.1–0.3%, while no repeated runs or error bars are given. The paper also omits hyperparameters, the split procedure, and the convergence threshold, and it cites unrelated references for baselines, making a fair comparison impossible to audit. These issues are distinct from the optimization-correctness question: the central empirical claim would fail if the differences are within noise, even if every update rule is correct. I therefore see the load-bearing concern as the absence of uncertainty quantification and reproducibility controls for the empirical comparison. The reader's verdict of CONDITIONAL remains appropriate, since the paper could be salvaged by adding seeds, error bars, hyperparameters, code, and corrected baseline descriptions; I would not move to REJECT or ACCEPT based on the current text.","tokens_in":11511,"tokens_out":4813,"duration_ms":46509,"concrete_test":"Run M1–M4 on the same D1–D4 splits with 10 random seeds or initializations, holding hyperparameters fixed after a joint tuning protocol, and report mean ± standard deviation plus paired bootstrap or permutation confidence intervals for the difference M4 − M3 (and M4 − M2). If the 95% CI for the D1/D2 MAE difference includes zero, or M4 does not win on a majority of seeds, the 'consistently outperforms' claim fails. As a prerequisite, the authors should release code and exact hyperparameters so the runs are reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that NSFT 'consistently outperforms' CP-SGD, NNCP, and CTF. For that claim to hold, the measured differences in Table III must be larger than run-to-run and configuration variability. They are not demonstrated to be. The largest gains are on D3/D4 (MAE about 1.5%, RMSE about 2.0–2.7% vs M3), but on D1 and D2 the margins over M3 are 0.10–0.15% MAE and 0.15–0.28% RMSE. The paper reports one run per method, no standard deviations, no seeds, no repeated initializations, no paired significance tests, and no hyperparameter values or selection procedure for any method, including R, F, the regularization coefficients, learning rates, and iteration count. Convergence is defined by a validation threshold whose value is missing ('drops below.'). The split construction is not specified (random versus temporal), so it is unclear whether the comparison tests temporal generalization at all. The garbled update equations in Eq. (12) and mismatched baseline citations [28],[34] add reproducibility problems, but the decisive vulnerability is that even a correct implementation of NSFT could easily sit within noise of M3 on the reported metrics. Without variance estimates or a paired comparison, the empirical claim is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a Non-negative Snowflake Factorization of Tensors (NSFT) model for predicting missing QoS values in user-service-time tensors. The model introduces a 'snowflake core tensor' to capture latent interactions and learns parameters via non-negative multiplicative updates (SLF-NMUT), including bias terms and Tikhonov regularization. Experiments on two WS-DREAM datasets (response time and throughput) with four train/validation/test splits compare NSFT against three tensor factorization baselines (CP-SGD, NNCP, CTF) and report lower MAE and RMSE for NSFT on all splits. The paper claims that NSFT consistently outperforms the baselines and offers better interpretability and generalization for dynamic QoS prediction.","tokens_in":11760,"tokens_out":3564,"duration_ms":34158,"significance":"If the reported empirical claim were substantiated, the contribution would be a modest but useful addition to the literature on non-negative tensor factorization for QoS prediction. The paper correctly evaluates on held-out test splits, so the central comparison is not circular. The authors also address a practically relevant problem and provide a schematic model design. However, the significance is limited by the absence of variance estimates, the tiny reported advantages over one baseline on two splits, and the lack of reproducible algorithmic details. No code, no convergence analysis, and no statistical tests are provided, so the current evidence does not establish that NSFT is reliably better than a strong baseline such as CTF.","major_comments":[{"comment":"The empirical claim that NSFT 'consistently outperforms' the baselines rests on differences as small as 0.1% (e.g., D1 MAE: M3 1.4331 vs M4 1.4315; D2 RMSE: M3 3.0789 vs M4 3.0704). The paper reports a single run per method with no standard deviations, seeds, repeated initializations, or paired significance tests, and it does not state how hyperparameters (ranks R and F, regularization coefficients, learning rates, iteration counts) were selected. Run-to-run or configuration variability of this magnitude could reverse the ranking; please provide variance estimates or paired comparisons and full hyperparameter settings for all methods.","section":"Section 4.3, Table III"},{"comment":"The multiplicative update rules are not presented in a verifiable form. The definition of β in Eq. (12) contains mixed subscripts and superscripts and is garbled; the relationship between the learning rates in Eq. (11) and the cancellation of 'negative components' is asserted rather than derived; and no monotonicity or convergence guarantee is given for the regularized square loss with the snowflake core. Since these updates are the basis of SLF-NMUT and of the experimental results, the derivation must be rewritten with all index ranges and definitions, and ideally with a convergence proof or a complete reference to one.","section":"Section 3.2, Eqs. (10)-(12)"},{"comment":"The train/validation/test splits D1-D4 are defined only by the ratio (1:2:7 or 2:2:6). It is not stated whether the split is random or along the time dimension, so it is unclear whether the evaluation tests temporal generalization, which is central to a 'dynamic QoS' model. Also, the convergence criterion in Section 4.2 is incomplete: the text says 'the difference in validation errors between successive iterations drops below.' with no threshold value. Please specify the splitting scheme, the convergence threshold, and all hyperparameter values.","section":"Section 4.1, 4.2"},{"comment":"The baseline citations do not match the named methods: reference [34] is 'Biased multiobjective optimization and decomposition algorithm', [28] is a PID-incorporated latent factorization paper, and [5] is 'Outlier-resilient web service QoS prediction'; none is clearly CP-SGD, NNCP, or CTF. Without correct references or implementation descriptions, the comparison cannot be reproduced. Please provide accurate citations and implementation details for each baseline.","section":"Section 4.3, baseline descriptions"}],"minor_comments":[{"comment":"Equation (5) is referenced in the text but not displayed; the numbering jumps from (4) to (6). Please renumber and ensure all referenced equations are present.","section":"Section 3.1, Eq. numbering"},{"comment":"The notation for the snowflake core tensor is not defined symbolically in a separate equation; the τ and γ sub/superscripts in Eq. (2) are unexplained. A clear definition of all symbols and index ranges is needed.","section":"Section 3.1, Eq. (2)"},{"comment":"Only the learning rates η_g and η_i are specified; the learning rate η_ir for the latent factor U is never defined, although the update for u_ir is given in Eq. (12).","section":"Section 3.2, Eq. (11)"},{"comment":"The text states that 'QoS entries equating to zero are excluded from the assessment' but does not explain how zeros are handled in the observed set Λ. Please clarify whether zero is treated as missing or as a valid observed value.","section":"Section 4.1, Table I"},{"comment":"The conclusion claims reduced computational complexity and training costs, but no complexity analysis or runtime measurements are presented anywhere in the paper.","section":"Section 5, Conclusion"},{"comment":"Several references are duplicated in the list (e.g., [25] and [40] are the same paper; [36] and [44] are the same paper). Please deduplicate and renumber.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on self-citations to a small group of authors, including several references that do not clearly correspond to the cited methods or claims; this should be checked for proper scholarly attribution. A revision that adds careful statistical validation, complete algorithmic details, and proper baseline references would substantially improve the paper. I do not see grounds for rejection if the authors can supply the missing reproducibility evidence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The paper's actual novelty is modest: a 'snowflake core tensor' that couples paired latent factors in a Tucker-like non-negative factorization, with a multiplicative update scheme. That is a real structural variation, not a restatement of an earlier model. And the empirical table is internally consistent: NSFT has lower MAE and RMSE than CP-SGD, NNCP, and CTF on all four WS-DREAM splits. So the paper is not empty.\n\nWhat it does well: it takes the non-negative tensor factorization line from Luo/Wu and adds a concrete core structure plus bias terms, and it evaluates on a standard benchmark with a usable train/valid/test design, though the split type is never specified. The paper also gives update rules, which is more than some papers in this area do.\n\nWhere it falls short, in order of severity. First, the central claim is 'consistently outperforms,' but the margins over the strongest baseline (CTF) on D1 and D2 are 0.1–0.3%. There are no error bars, no seeds, no repeated runs, no paired tests. One run per method cannot distinguish a real improvement from initialization noise. The conclusion also claims robustness under 'varying sparsity levels,' but all four datasets are fixed at 74.06% density; no sparsity sweep exists. Second, the math is not checkable as printed. Eq. (12) has tangled indices, the learning rate setting is described as canceling negative terms rather than derived, and the convergence threshold is literally 'drops below.' with no value. Hyperparameters R, F, lambda, and iteration counts are all missing. Third, some baseline citations look mismatched (e.g., [34] is a biased multiobjective optimization paper, not CP-SGD). That is sloppy rather than fatal, but it compounds the reproducibility problem.\n\nThe reader's weakest assumption is roughly right: if the update rules are wrong or unstable, the numbers collapse. But I would not call the claim circular; held-out test splits make the evaluation forward-looking. The problem is uncertainty, not circularity.\n\nWho this is for: someone working on non-negative tensor factorization for QoS might cite this as a known variant, but the current text does not give them enough to reproduce or build on. I would not bring it to a reading group.\n\nRecommendation: send it to peer review, because the model is concrete and the empirical claim is falsifiable. Tell the authors that code, exact hyperparameters, variance estimates, a corrected derivation, and a real sparsity experiment are all required before acceptance. As is, it is a conditional: a plausible idea, not yet supported evidence.","headline":"A plausible tensor-factorization variant whose empirical claim is too thinly supported by 0.1% margins and no error bars, but concrete enough to warrant referee attention with a demand for code and variance estimates.","tokens_in":12317,"tokens_out":4146,"would_cite":false,"duration_ms":40334,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A snowflake-structured tensor core, trained with nonnegative multiplicative updates, predicts missing QoS values with lower MAE and RMSE than CP-SGD, NNCP, and CTF on all four WS-DREAM splits.","keywords":["QoS prediction","tensor factorization","non-negative constraint","multiplicative update","snowflake core tensor","tensor completion","WS-DREAM","dynamic QoS data"],"falsifier":"Implement the printed update equations exactly and run the four WS-DREAM splits with the same train/validation/test masks; if validation loss does not decrease monotonically or test MAE and RMSE deviate beyond a small tolerance from Table III, the central claim fails. A cheaper check is a synthetic tensor with known factors: if the snowflake-core objective does not decrease at every iteration, the update rules are not minimizing the stated loss.","tokens_in":11268,"feed_emoji":"❄️","tokens_out":8531,"duration_ms":75935,"temperature":0.7,"pith_summary":"The paper proposes the Non-negative Snowflake Factorization of Tensors (NSFT) model for predicting missing quality-of-service (QoS) values in sparse user-service-time tensors. The central idea is to give the factorization a snowflake-shaped core tensor, a set of nonnegative coupling weights that tie user, service, and time latent factors together in several rank-one products per component, and to learn all parameters with a multiplicative update rule that preserves non-negativity. On WS-DREAM response-time and throughput data split into four configurations, NSFT reports lower mean absolute error and lower root mean square error than the CP-SGD, NNCP, and CTF baselines. If the claim holds, service recommendation systems can fill in unobserved QoS entries more accurately from historical logs, without additional data collection.","feed_headline":"Snowflake tensor core cuts QoS prediction error across four splits","feed_subtitle":"Nonnegative multiplicative updates beat CP-SGD, NNCP, and CTF on WS-DREAM response time and throughput data.","key_machinery":"The central object is the snowflake core tensor: a structured set of nonnegative coupling tensors that, instead of one shared core, provides several rank-one interaction terms per latent component to connect the user, service, and time modes. It is paired with the SLF-NMUT learning rule, a single-latent-factor nonnegative multiplicative update in which the learning rates are set so that negative gradient contributions cancel, keeping all latent factors and core entries nonnegative. This machinery lets the model fit sparse QoS tensors with a Euclidean loss while preserving the nonnegativity of response time and throughput data.","core_discovery":"The paper claims that a tensor factorization with a snowflake-structured core learns dynamic user-service-time patterns better than conventional low-rank tensor completion. In the model, each latent component contributes several nonnegative products of user, service, and time latent vectors, with the interaction weights stored in the snowflake core tensor; linear bias vectors for users, services, and time slices are added to absorb level shifts. Parameters are fit by minimizing a Tikhonov-regularized Euclidean loss, and the SLF-NMUT scheme converts gradient descent into nonnegative multiplicative updates by choosing learning rates that cancel negative gradient terms. On the WS-DREAM datasets, the authors report that NSFT achieves lower MAE and RMSE than CP-SGD, NNCP, and CTF on all four splits, for example an MAE of 1.4315 versus 1.4331 for the closest baseline on D1, and interpret this as evidence that the snowflake core adds useful modeling capacity for incomplete dynamic QoS tensors.","pith_inferences":["The reported improvements over NNCP and CTF are small on most splits, so without repeated-run standard deviations or significance tests the 'consistently outperforms' claim may be within noise; a multi-seed replication would settle that.","The snowflake core is an architectural choice independent of QoS, so applying the same factorization to traffic recovery, network link prediction, or recommender tensors would test whether the structure, rather than the dataset, drives the accuracy gain.","An ablation that replaces the snowflake core with a single diagonal core while keeping the SLF-NMUT updates would isolate how much of the gain comes from the core shape versus the update rule.","The paper gives no explicit values for rank R, factor count F, or regularization coefficients, so releasing those settings is the minimal step needed to turn the reported numbers into a reproducible baseline for later work."],"forward_implications":["On WS-DREAM, NSFT produces lower MAE and RMSE than CP-SGD, NNCP, and CTF across all four train/validation/test splits, so it can be used as a drop-in completion model for response-time and throughput data.","Because the model maintains nonnegativity through multiplicative updates, it can be trained on the large, sparse QoS tensors without resorting to projected-gradient or ADMM solvers.","The reported robustness under increasing missing-entry ratios implies NSFT stays accurate when the observed fraction of the tensor is low.","The snowflake core enriches the interaction structure of a low-rank completion model, which the authors take as evidence that such structured cores improve dynamic QoS prediction."],"supporting_citations":[{"why":"Supplies the WS-DREAM response-time and throughput datasets used for all experiments and the claim's empirical basis.","marker":"[35]"},{"why":"Defines the CP-SGD baseline (M1) whose MAE and RMSE NSFT must beat.","marker":"[34]"},{"why":"Defines the NNCP baseline (M2) and motivates the nonnegative multiplicative update direction.","marker":"[28]"},{"why":"Defines the CTF baseline (M3) and the outlier-resilient QoS prediction setting.","marker":"[5]"},{"why":"Provides the biased nonnegative latent factorization of tensors approach on which the temporal QoS model is built.","marker":"[4]"},{"why":"Supplies the biased nonnegative Tucker factorization of tensors that the snowflake core extends.","marker":"[18]"},{"why":"Supplies diversified regularization schemes for nonnegative tensor factorization that the objective borrows.","marker":"[19]"}],"fun_headline_variants":["Snowflake tensor core improves QoS prediction","Nonnegative snowflake factorization predicts QoS better","Snowflake core reduces QoS error in four splits","Tensor snowflake beats baselines for QoS prediction","Snowflake tensor factorization lifts QoS accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's accuracy claim rests on the multiplicative update rules in Section 3.2 actually driving the regularized loss down while keeping every factor nonnegative; if those update equations are not correct or stable, the reported MAE and RMSE values would not be reproducible.","fun_headline_variants_meta":{"raw":{"variants":["Snowflake tensor core improves QoS prediction","Nonnegative snowflake factorization predicts QoS better","Snowflake core reduces QoS error in four splits","Tensor snowflake beats baselines for QoS prediction","Snowflake tensor factorization lifts QoS accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000395,"raw_usage":{"total_tokens":2031,"prompt_tokens":867,"completion_tokens":1164,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":483,"completion_tokens_details":{"reasoning_tokens":1098}},"tokens_in":483,"tokens_out":1164,"duration_ms":7562,"temperature":1.0,"reasoning_tokens":1098,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:45:47.550050+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Implement the printed update equations exactly and run the four WS-DREAM splits with the same train/validation/test masks; if validation loss does not decrease monotonically or test MAE and RMSE deviate beyond a small tolerance from Table III, the central claim fails. A cheaper check is a synthetic tensor with known factors: if the snowflake-core objective does not decrease at every iteration, the update rules are not minimizing the stated loss.","supporting_citations":[{"cited_title":"Web service QoS prediction using improved software source code metrics,","cited_arxiv_id":null,"evidence_quote":"Supplies the WS-DREAM response-time and throughput datasets used for all experiments and the claim's empirical basis."},{"cited_title":"Biased multiobjective optimization and decomposition algorithm,","cited_arxiv_id":null,"evidence_quote":"Defines the CP-SGD baseline (M1) whose MAE and RMSE NSFT must beat."},{"cited_title":"A PID -incorporated latent factorization of tensors approach to dynamically weighted directed network analysis,","cited_arxiv_id":null,"evidence_quote":"Defines the NNCP baseline (M2) and motivates the nonnegative multiplicative update direction."},{"cited_title":"Outlier-resilient web service QoS prediction,","cited_arxiv_id":null,"evidence_quote":"Defines the CTF baseline (M3) and the outlier-resilient QoS prediction setting."},{"cited_title":"Temporal pattern-aware QoS prediction via biased non-negative latent factorization of tensors,","cited_arxiv_id":null,"evidence_quote":"Provides the biased nonnegative latent factorization of tensors approach on which the temporal QoS model is built."},{"cited_title":"Temporal pattern-aware QoS prediction by biased non-negative Tucker factorization of tensors,","cited_arxiv_id":null,"evidence_quote":"Supplies the biased nonnegative Tucker factorization of tensors that the snowflake core extends."},{"cited_title":"Advancing non -negative latent factorization of tensors with diversified regularization schemes,","cited_arxiv_id":null,"evidence_quote":"Supplies diversified regularization schemes for nonnegative tensor factorization that the objective borrows."}],"review_version":1}