{"id":"003d84d1-bf9b-4aac-aac8-3c1d77afc1e1","arxiv_id":"2504.14565","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"Recommender embeddings built from multi-view, dynamically pruned cluster centers give modest RMSE gains over classic matrix factorization baselines.","lead":"This paper proposes MFDMC, a matrix factorization model for recommendations in which user and item embeddings are built as weighted combinations of cluster centers across several views. It reports small RMSE gains over classic baselines on six datasets and argues the resulting representations are lower-dimensional and more interpretable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The balancing property claimed for Eq. 7 is arithmetically false: uniform weights give entropy loss t/e, not a constant, so the stated mechanism for stable multi-view optimization is unsupported.","rationale":"The reader's weakest-assumption analysis correctly identifies Eq. 7 as a critical point, and the arithmetic is right: uniform weights yield loss t/e, not a t-invariant constant. This is a genuine, concrete flaw in the paper's stated mechanism. It does not, however, automatically invalidate the empirical observation that MFDMC achieved lower RMSE in the reported tables; that could still be true for reasons other than the claimed balancing property, such as the extra clustering losses or the particular hyperparameter tuning. The appropriate response is therefore to require a correction and a verification, which is exactly what a CONDITIONAL verdict expresses. I considered whether a more severe verdict is warranted because the model's behavior may differ from the paper's description, but the paper is a preprint and the central construction is coherent enough that a corrected normalization or a revised explanation could plausibly salvage the claim. No other concern seems more load-bearing: the baselines are weak and no uncertainties are reported, but those are standard limitations that do not undercut the central mechanism as sharply as the Eq. 7 failure does. The reader's verdict should remain CONDITIONAL, and no further adjustment is needed.","tokens_in":11243,"tokens_out":5977,"duration_ms":59265,"concrete_test":"Analytically recompute Eq. 8 under Eq. 7 for uniform weights with t=3 and t=10: uniform w' = 1/t maps to w'' = 1/e for each center, giving loss_2 = t/e, i.e., 3/e and 10/e respectively. If the two values are unequal, the claimed invariance is false; if the authors' code instead renormalizes w'' per view before computing entropy, Eq. 7 should be revised to include that normalization and the RMSE comparison rerun to confirm the reported gains persist.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that dynamic multi-view clustering lets matrix factorization use the representation space more effectively, and a key load-bearing component of that argument is the mapping function in Eq. 7, which is said to ensure that the worst-case uniform weight distribution has the same loss regardless of the number of cluster centers t in a view. This invariance is the only stated basis for balanced optimization across views with different t. Substituting a uniform softmax output, w' = 1/t for each of the t centers, into Eq. 7 gives w'' = 1/e for every center, because the first branch evaluates to (t/e)*(1/t) = 1/e and the second branch also gives 1/e at the boundary. The entropy-based loss in Eq. 8 is then -sum_i (1/e) log(1/e) = t/e, which grows linearly with t. Thus a view with 10 centers has 10/3 times the loss of a view with 3 centers under the same uniform distribution, directly contradicting the paper's claim that the mapping makes the uniform distribution loss consistent across views. This is not a cosmetic issue: the stability of training, the interpretation of the piecewise mapping in Fig. 2, and the attribution of the method's gains to balanced multi-view optimization all rely on this invariance. As written, the method's explanatory mechanism is unsupported, even though the empirical RMSE tables could in principle still be reproducible through other effects. The paper therefore needs either a corrected mapping, a renormalization step, or a revised explanation before the central claim can be accepted as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MFDMC, a matrix-factorization recommender model that learns user and item representations as weighted combinations of cluster centers across multiple views, with dynamic pruning of clusters and an entropy-based regularizer on the weights. The authors report RMSE improvements over FunkMF, BiasedMF, PMF, SVD, and Glocal-K on six datasets, an ablation study on MovieLens-1M, and a small synthetic computer-vision experiment to illustrate interpretability and transferability. The core claimed contribution is that dynamic multi-view clustering makes better use of the latent representation space, yielding more accurate and more interpretable recommendations at low dimensionality.","tokens_in":11634,"tokens_out":3907,"duration_ms":37002,"significance":"If the method and experiments were fully valid, the idea of coupling matrix factorization with dynamic multi-view clustering would be a useful and timely contribution to representation learning for recommender systems. The paper's strengths include evaluation on multiple public datasets, an ablation study, a visualization-based interpretability analysis, and an extension to a downstream image-classification task. However, the paper as written contains a demonstrable arithmetic error in the central balancing mechanism, incomplete formal specification of the model, and an experimental section without variance estimates or modern baselines. These issues currently prevent the claims from being accepted at face value.","major_comments":[{"comment":"The claimed invariance of the uniform-distribution entropy loss to the number of cluster centers t is arithmetically false. Substituting w' = 1/t into Eq. (7) gives w'' = 1/e for every center, so Eq. (8) yields loss = t/e, which grows linearly with t. A view with 10 centers has 10/3 times the loss of a view with 3 centers under the same uniform distribution. Since this invariance is the stated motivation for the mapping function and the basis for balanced multi-view optimization, the explanatory mechanism is unsupported. Please correct the mapping, add a renormalization step, or revise the claim.","section":"Section III.C, Eqs. (7)-(8)"},{"comment":"The notation in Eq. (2) is internally inconsistent and prevents reproduction. The user index i is reused as the cluster-center index, the concatenation operator is not defined, and the summation bounds are unclear. Similarly, Eq. (8) sums w^{user''}_{i,j} over i = 0 to m, but Table I defines W as a v x e matrix without a per-user dimension. Please rewrite the model with distinct indices for users, views, and centers, and state the exact tensor shapes of all weight variables.","section":"Section III.A, Eq. (2)"},{"comment":"The central claim that MFDMC consistently outperforms competitors is supported only by single-point RMSE values with no error bars, repeated runs, or significance tests. The baselines are also limited to classic matrix-factorization methods; no recent deep MF, multi-view clustering, or graph-based recommender is included, despite the related-work section positioning the paper against such methods. Please add repeated-run statistics with variance reporting and compare against at least one or two modern strong baselines.","section":"Section IV.C, Tables III-IV"},{"comment":"The abstract and introduction make strong claims about substantial reductions in time and computational demands and about scalability to web-scale data, but the experiments contain no runtime, memory, or convergence measurements anywhere. Tables III and IV report only RMSE. Please provide an empirical complexity evaluation or a formal complexity analysis, or temper the computational-efficiency claims to what the experiments actually demonstrate.","section":"Sections I and IV"},{"comment":"The text states that \"centers in the same location in the view can be shared\" between users and items, but Eqs. (2)-(5) define separate C^{user} and C^{item} variables and do not reflect sharing. If shared centers are used in the reported experiments, the objective must be amended to show how sharing is realized; otherwise the distinction between \"Not Share\" and \"Share\" in Table V is unclear. Please specify which equations govern the shared-center configuration.","section":"Section IV.B and Table V"}],"minor_comments":[{"comment":"The summation notation in Eq. (4) is malformed: \"N,k∈Si,jX k=0\" is not a well-formed expression. Please rewrite it with explicit index ranges.","section":"Section III.B, Eq. (4)"},{"comment":"The notation \"e,v ∈ Z\" appears to be a typo for \"t,v ∈ Z\", and the constraint b = d/v requires that v divides d; please state this divisibility assumption explicitly.","section":"Table I"},{"comment":"The example comparing view1 with 3 centers and view2 with 10 centers appears to claim the opposite of the intended effect: under a uniform distribution, the total entropy loss should be smaller for fewer centers, not larger. Please clarify whether the text is describing per-center or total loss.","section":"Section III.C"},{"comment":"Please specify the axis along which softmax is applied: softmax over the centers within each view, or over all weights jointly.","section":"Section III.C, Eq. (6)"},{"comment":"The instruction \"W ← cluster-wise mean of user/item in W\" is ambiguous; please define precisely what is averaged and over which subset of users or items.","section":"Algorithm 1"},{"comment":"There are several typos, including \"FunckMF\" in the Introduction, \"RSME\" in Eq. (19), and \"MovieLen\" instead of \"MovieLens\" in Section IV.A. Also, references [39] and [42] appear to be the same paper (Explainable Matrix Factorization) and should be merged or disambiguated.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is a synthesis rather than a breakthrough, but a genuinely useful one. MFDMC integrates dynamic multi-view clustering into matrix factorization: each user/item latent vector is the concatenation of weighted sums of cluster centers across views, with spread, proximity, and entropy losses plus pruning. That combination is new as far as I can tell. And the headline numbers are plausible: MFDMC with 16 latent dimensions beats FunkMF and BiasedMF at 60 dimensions on MovieLens-1M/100k, with consistent improvements on the other four datasets. The ablation study is honest, and the synthetic CV sanity check is a nice touch even though it's toy data.\n\nThe soft spots are significant. The biggest is Eq. 7. The stress-test note is correct: for a uniform weight distribution, w'=1/t, the mapping gives w''=1/e for every center, so the entropy loss in Eq. 8 is t/e, which grows with t. The paper's claim that the mapping makes the uniform-distribution loss invariant across views is simply false. That's not a cosmetic error — the mapping is introduced specifically to balance optimization across views, and the paper's story about why MFDMC works relies on it. The empirical RMSE results might still be reproducible through the other losses, but the stated mechanism is unsupported as written. The authors need to either correct the mapping, renormalize properly, or reframe the explanation.\n\nOther issues are more about evidence. There are no error bars or significance tests, the baselines are classic MF methods rather than current state of the art, and no runtime or memory measurements are reported even though the paper claims web-scale efficiency. The CV experiment is a 9-class synthetic dataset, so the generalizability claim is overreaching. And there are some typos/notation inconsistencies (e.g., t vs e for the number of centers) that make parts of the math harder to follow than necessary.\n\nWho's this for? Someone working on interpretable or low-dimensional recommender embeddings would get value from the MFDMC idea, and the failure mode of Eq. 7 is instructive. I'd send it to a serious referee, but with the expectation of major revision. The architecture could be a decent contribution if the normalization issue is fixed and the experiments are tightened. As it stands, don't take the web-scale claims at face value.","headline":"MFDMC is a worthwhile synthesis of multi-view clustering and matrix factorization with plausible RMSE gains, but a false normalization claim in Eq. 7 undercuts the stated balancing mechanism and needs fixing.","tokens_in":12091,"tokens_out":2768,"would_cite":false,"duration_ms":23742,"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":"The paper claims that replacing free latent vectors with weighted, pruned cluster centers across views makes matrix factorization both more accurate and more interpretable.","keywords":["matrix factorization","multi-view clustering","recommender systems","representation learning","dynamic clustering","interpretability","latent space utilization","RMSE"],"falsifier":"Substitute uniform weights into Eq. 7 and Eq. 8 for two views with different numbers of centers: the resulting entropy losses are $t/e$ for $t$ centers, so the claimed invariance fails; a reader can verify this with a one-line calculation, and it directly undermines the stated purpose of the mapping function.","tokens_in":11065,"feed_emoji":"🎯","tokens_out":9073,"duration_ms":78168,"temperature":0.7,"pith_summary":"The paper claims that standard matrix factorization leaves much of its latent space unused, and that replacing free user and item vectors with weighted combinations of cluster centers organized into several views, learned end-to-end with clustering losses and dynamic pruning, remedies that waste. On six recommendation datasets it reports lower RMSE with a latent dimension of 16 than baselines achieve at dimension 60, including an improvement of about 0.025 RMSE over the FunkMF baseline on MovieLens-100k. It further reports that clusters take on interpretable meanings, such as movie genres, so a recommendation can be traced back to the clusters that dominate a user's representation. The same dynamic multi-view clustering module, attached to CNN features, is reported to transfer to image classification on a synthetic shape-and-color dataset. At stake is whether recommenders can be smaller, faster, and explainable without losing accuracy.","feed_headline":"Dynamic clusters lower recommendation error at smaller dimensions","feed_subtitle":"A clustered factorization model reports RMSE gains on six datasets using fewer latent dimensions than its baselines.","key_machinery":"The central object is the multi-view weighted-cluster representation: each user or item latent vector is formed by concatenating, over $v$ views, the weighted sum of $t$ cluster centers, with the user form written as $\\mathbf{p}_i = \\bigoplus_{j=0}^{v} \\sum_{i'=0}^{t} \\mathbf{c}^{\\mathrm{user}}_{i',j}\\, w^{\\mathrm{user}}_{i',j}$ and an analogous formula for items. Cluster centers are learned with a spread loss and a proximity MSE, while the weights are softmax-normalized and then reweighted by the mapping function of Eq. 7 so that entropy losses are claimed to be comparable across views. A pruning step, run every $I_p$ epochs, removes centers whose mean weight falls below a threshold, so the number of centers per view changes during training. This construction carries the argument: it is the mechanism intended to use the representation space fully, produce interpretable clusters, and keep optimization balanced across views.","core_discovery":"MFDMC's central discovery is that a matrix-factorization model can be built so that every user and item vector is a concatenation of weighted sums of cluster centers across several views, and that this construction, optimized with a spread loss that keeps centers apart, a proximity loss that binds points to centers, an entropy loss on the weights, and a pruning rule that deletes low-weight centers, makes better use of the representation space than unconstrained latent vectors. On the paper's six datasets, MFDMC with latent dimension 16 reports RMSE below FunkMF(60), BiasedMF(60), PMF(60), SVD, and Glocal-K; on MovieLens-100k the gap to FunkMF is about 0.025 RMSE. The authors also report that the learned clusters correspond to semantic categories, and that the same module transfers to image classification.","pith_inferences":["Beyond the paper: if the Eq.-7 balancing property fails, the reported gains may be carried by the spread and proximity losses instead, which a simple ablation that removes only the mapping function could isolate.","Beyond the paper: dynamic pruning currently only removes weak centers; adding a splitting operation would let the model track emerging item categories or user interests in streaming data.","Beyond the paper: the interpretability claim implies a testable behavior, namely that users whose vectors are dominated by a high-rated animation cluster should receive animation-heavy recommendations, which could be checked on held-out interactions.","Beyond the paper: the same weighted-cluster representation could be applied to other embedding-based models, such as graph embeddings, where a node's multiple roles could be captured by separate views."],"forward_implications":["At latent dimension 16, MFDMC reports lower RMSE than baselines run at dimension 60, so recommender models could run with far fewer parameters for the same accuracy.","Because clustering happens inside the same training loop, the separate clustering stage of two-stage pipelines disappears, reducing computational cost at web scale.","User and item embeddings become sums of view-specific cluster centers, so a recommendation can be explained by which clusters, for example genre clusters, dominate the user's vector.","The same dynamic clustering module, attached to CNN features, transfers to image classification, suggesting the mechanism is not specific to ratings data.","Dynamic pruning means the number of clusters is not fixed in advance; views can shrink during training, adapting to the data rather than requiring a grid search over cluster counts."],"supporting_citations":[{"why":"This is the FunkMF baseline objective; MFDMC is compared directly against it on MovieLens-100k and MovieLens-1M.","marker":"[30]"},{"why":"This supplies the BiasedMF baseline, which MFDMC must beat on all six datasets.","marker":"[8]"},{"why":"This supplies the probabilistic matrix factorization baseline used in the comparison tables.","marker":"[31]"},{"why":"This supplies the Glocal-K baseline, another matrix-factorization competitor in the tables.","marker":"[52]"},{"why":"This provides the MovieLens-100k and MovieLens-1M datasets on which the headline RMSE improvement is measured.","marker":"[45]"},{"why":"This provides the Amazon-video dataset used in the large-scale evaluation.","marker":"[46]"},{"why":"This provides the Books-across dataset used in the large-scale evaluation.","marker":"[47]"},{"why":"This supplies the hyperparameter optimization tool used in the implementation of all experiments.","marker":"[49]"}],"fun_headline_variants":["Dynamic clusters cut recommendation error with fewer dimensions","End-to-end clustering model shrinks latent space, improves RMSE","MFDMC: Adaptive multi-view clustering for leaner recommendations","Dynamic multi-view clusters beat larger latent models","Unified clustering and factorization outperforms larger latent models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the reweighting formula in Eq. 7 makes the worst-case entropy loss independent of the number of clusters in a view; substituting uniform weights contradicts this because the loss grows with the cluster count.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic clusters cut recommendation error with fewer dimensions","End-to-end clustering model shrinks latent space, improves RMSE","MFDMC: Adaptive multi-view clustering for leaner recommendations","Dynamic multi-view clusters beat larger latent models","Unified clustering and factorization outperforms larger latent models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000451,"raw_usage":{"total_tokens":2243,"prompt_tokens":888,"completion_tokens":1355,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":1278}},"tokens_in":504,"tokens_out":1355,"duration_ms":11829,"temperature":1.0,"reasoning_tokens":1278,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:46:02.522233+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Substitute uniform weights into Eq. 7 and Eq. 8 for two views with different numbers of centers: the resulting entropy losses are $t/e$ for $t$ centers, so the claimed invariance fails; a reader can verify this with a one-line calculation, and it directly undermines the stated purpose of the mapping function.","supporting_citations":[{"cited_title":"Probabilistic matrix factorization,","cited_arxiv_id":null,"evidence_quote":"This supplies the probabilistic matrix factorization baseline used in the comparison tables."},{"cited_title":"Glocal-k: Global and local kernels for recommender systems,","cited_arxiv_id":null,"evidence_quote":"This supplies the Glocal-K baseline, another matrix-factorization competitor in the tables."},{"cited_title":"Ups and downs: Modeling the visual evolution of fashion trends with one-class collaborative filtering,","cited_arxiv_id":null,"evidence_quote":"This provides the Amazon-video dataset used in the large-scale evaluation."},{"cited_title":"Improving recommendation lists through topic diversification,","cited_arxiv_id":null,"evidence_quote":"This provides the Books-across dataset used in the large-scale evaluation."}],"review_version":1}