{"id":"52689f2b-27eb-496a-865c-8f152e93c4bf","arxiv_id":"2501.18997","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CDiff4Rec improves diffusion recommenders by injecting item-content pseudo-users and real-user neighbor predictions into the denoising objective, beating DiffRec and other baselines on Yelp, Amazon-Game, and Citeulike-t.","lead":"This paper adds a diffusion-based recommender that turns item review words into virtual 'pseudo-users' and blends the predictions of a user's most similar real and pseudo neighbors into the denoised preference vector. The approach is tested on three public datasets and improves ranking metrics over a standard diffusion recommender, at small extra compute cost.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 6 is not a well-defined training objective as written: neighbor and pseudo-user predictions require their own noisy inputs, but Eq. 8 samples only the query user's x_t; the claimed gains in Table 2 depend on an unspecified inference/training protocol.","rationale":"I read the paper as making a concrete empirical claim: incorporating denoised predictions from real and pseudo neighbors into the diffusion reconstruction loss improves top-N accuracy, and Table 2 is the evidence. For that claim to be true, Eq. (6) must denote an actual computable quantity. The most load-bearing weakness is that the neighbor terms are never given a sampling or inference definition: the model's only denoiser input is x_t, but Eq. (8) only defines x_t for the query user, leaving the neighbor and pseudo-user forward processes unspecified. This is an internal consistency issue, not merely a disagreement with current practice. The second issue about continuous TF-IDF pseudo-users being fed to a denoiser trained on binary vectors is independently load-bearing, because the pseudo-user ablation shows substantial gains; the paper never states a loss that teaches the model to denoise these vectors. Both issues are fixable by adding detail, so the correct disposition is to require those details before accepting the result. The reader's conditional verdict already captures this, and their weakest assumption points to the same Eq. (6) protocol gap, so I agree with the assessment. There is no machine-checked proof or released code that could independently verify the construction, which increases the need for a precise protocol specification.","tokens_in":8808,"tokens_out":3996,"duration_ms":42598,"concrete_test":"Write out an explicit sampling protocol for Eq. (6): for each query user u at timestep t, state how x_t^{(ru_i)} and x_t^{(pu_j)} are drawn (e.g., independently from q(x_t | r_{ru_i}) and q(x_t | m_{pu_j})), then rewrite Eq. (8) as an expectation over all of these variables and derive the actual gradient used. Implement that protocol or the authors' intended one, rerun Table 2, and ablate by replacing neighbor and pseudo-user predictions with their clean (noise-free) vectors while keeping the same attention scores; if the margins and significance levels do not survive, the reported gains depend on an unstated denoising procedure for neighbors.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that blending the query user's denoised vector with denoised real and pseudo neighbors produces the consistent gains in Table 2. That claim rests entirely on Eq. (6), but Eq. (6) is underspecified to the point of being non-computable as stated. The denoiser x̂_θ takes a noisy input x_t and timestep t; therefore the neighbor terms ˆr_{ru_i} and ˆm_{pu_j} must be computed as x̂_θ(x_t^{(ru_i)}, t) and x̂_θ(x_t^{(pu_j)}, t), using each neighbor's own noisy vector. The training objective Eq. (8) is written as E_{q(x_t|x0)}[ C ||ˆr'_u - r_u||^2 ] with the expectation only over the query user's forward process. No sampling distribution is specified for the neighbors' noisy vectors, and no joint expectation appears. If neighbors are sampled independently, Eq. (8) is incomplete; if their noisy vectors are reused from the query's x_t, that is a different and unstated model. At inference, the paper says the query user's corrupted input is iteratively denoised, but producing each neighbor prediction would require separately running the diffusion chain for every neighbor or another unspecified approximation. This is not cosmetic: the headline improvements (Yelp R@20 0.1145 vs. 0.1099 with p≤0.0005; AM-Game 0.2255 vs. 0.2193; Citeulike-t 0.1616 vs. 0.1591) are measured under this protocol. A second, closely related gap compounds the problem: pseudo-user vectors m_p are continuous TF-IDF values in [0,1], while the denoiser is explicitly trained on binary user vectors, and Eq. (8) contains no pseudo-user reconstruction loss. Since Table 2 shows Ours (+Pseudo-users) alone already beats DiffRec and the full model on two datasets, the model must somehow denoise these continuous inputs, but no training signal for that is defined.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CDiff4Rec, a diffusion-based collaborative filtering method that augments the query user's denoised interaction vector with the denoised vectors of two kinds of neighbors: top-K real users selected by cosine behavior similarity, and top-K pseudo-users constructed from review words by treating each word as a pseudo-user with TF-IDF-normalized interaction weights. The blended prediction is a convex combination with attention scores, and the training loss extends the DiffRec objective by replacing the denoised reconstruction with this blended vector. The paper reports Recall@20 and NDCG@20 gains over BPRMF, LightGCN, Ease, MultiVAE, ConVAE, and DiffRec on Yelp, Amazon-Game, and Citeulike-t, with ablations showing that both real and pseudo neighbors contribute, plus an accuracy-efficiency comparison against DiffRec.","tokens_in":9220,"tokens_out":6240,"duration_ms":63099,"significance":"If the method is fully specified and reproducible, the paper makes a useful contribution: it shows that item-content signals and collaborative neighbor signals can be folded directly into the diffusion objective without changing the backbone, and the pseudo-user construction is a simple and portable way to exploit review texts in a CF-oriented diffusion model. The consistent gains across three datasets, the component ablations, and the modest wall-clock overhead in Table 3 are genuine strengths. However, the central equations as written omit the sampling and inference protocol on which the headline gains depend, and the treatment of continuous pseudo-user vectors with a binary-trained denoiser is an unstated structural assumption, so the empirical claims cannot currently be checked from the text.","major_comments":[{"comment":"As written, the training objective is not computable: x̂_θ requires a noisy input and a timestep, so the neighbor terms ˆr_{ru_i} and ˆm_{pu_j} in Eq. (6) each require their own noisy vectors x_t^{(ru_i)} and x_t^{(pu_j)}, yet Eq. (8) samples only q(x_t | x0) for the query user and specifies no joint distribution over the neighbors' noisy inputs. Please state the sampling scheme used for neighbor inputs during training, or, if the query's x_t is reused for all neighbors, say so explicitly and justify this choice.","section":"§3.3, Eq. (6), Eq. (8)"},{"comment":"The inference protocol for Eq. (6) is unspecified: computing each neighbor term requires running the denoiser on a noisy input for every neighbor, and the paper does not say whether this is done with a separate reverse chain per neighbor, with a shared noisy input, or with cached forward predictions. This matters for the wall-clock comparison in Table 3 and for the reproducibility of the gains in Table 2.","section":"§3.3, §4.1.3"},{"comment":"The denoiser is trained on binary user vectors through Eq. (8), but pseudo-user vectors m_{pu} are continuous TF-IDF values normalized to [0,1]. The paper therefore relies on the unstated assumption that a denoiser trained on binary inputs generalizes to continuous pseudo-user inputs; please add a pseudo-user training term or provide evidence (e.g., a validation experiment) that this transfer works.","section":"§3.1, §3.3"},{"comment":"As written, the top-K real-neighbor set U_u^{ru} is selected from all users including the query user u itself. Because cosine distance between r_u and itself is zero, the query user would typically be its own nearest neighbor and would enter the blend in Eq. (6), potentially inflating the reported improvements. Please exclude the query user from the neighbor search or clarify if self-neighbors are intentionally retained.","section":"§3.2, Eq. (5)"},{"comment":"The definition of \\bar\\alpha_t as a product over t'=1 to T makes \\bar\\alpha_t independent of t, which would break the forward process in Eq. (1) and the mean and loss formulas in Eqs. (2) and (4). This should be a product up to t; please correct the index and verify that the implementation uses the corrected definition.","section":"§2, Eq. (1)"},{"comment":"The actual values of α, β, γ, K, and |F| used for the reported main results are missing; the text gives only search ranges, and Table 4 shows non-monotonic sensitivity to K and |F|. In addition, the paired t-test is not described (unit of analysis, number of paired observations, and whether it is over users or repeated runs). Please report the selected hyperparameters and the test details so the significance claims can be checked.","section":"§4.1.3, Table 2"}],"minor_comments":[{"comment":"The argsort notation does not state whether the order is ascending or descending; since distance is being sorted, please define the intended ordering explicitly.","section":"§3.2, Eq. (5)"},{"comment":"Table 4 does not state the reported metric in the caption; it should be identified (presumably R@20), and the caption of Table 3 should also spell out that the metrics are Recall and NDCG at the listed ranks.","section":"§4.3.2, Table 4"},{"comment":"The sentence claiming that top-20 neighbors and 1,000 pseudo-users are superior across the three datasets is not fully supported by Table 4, where Yelp achieves its best R@20 in the K=50 column for the 1,000-pseudo-user setting; please clarify the selection criterion or qualify the claim.","section":"§4.3.2"},{"comment":"The conclusion repeats the abstract's phrase \"loss of personalized information\" as if it were a quantity; consider rephrasing to \"mitigates the loss of personalized information\" or \"reduces the degradation of personalized information\" for clarity.","section":"§5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a compact WWW Companion paper. The core idea is interesting and the empirical gains are consistent, but the main method is underspecified exactly where the headline comparisons depend on it. I see no evidence of misconduct; the one same-group self-citation (Han et al.) is incidental and does not affect the claims. My main concern is that the reported results cannot be reproduced or fully trusted until the training and inference protocol for Eq. (6), the handling of continuous pseudo-user vectors, and the hyperparameter choices are clarified. I recommend major revision rather than rejection because these gaps appear fixable with additional specification and targeted experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read: the paper has a genuinely useful idea -- treating review words as pseudo-users and adding neighbor denoised predictions into the diffusion objective -- and the reported gains over DiffRec/Ease are consistent across three datasets. That part earns its keep. But the formal description is incomplete: Eq. (6) and Eq. (8) do not define a computable objective unless the authors specify how neighbor predictions are generated during training and inference.\n\nWhat's new: this combination of pseudo-users from item features and neighbor aggregation inside the denoising target does not appear in the cited diffusion-recommender literature. The individual tricks are standard elsewhere (review words as users, kNN smoothing), but the integration is a legitimate extension of DiffRec. The ablations make sense: real neighbors alone add a bit, pseudo-users add a bit, and the combined model is best on all three datasets. Wall time is reported to increase only a little, which is credible if the neighbor sets are cached.\n\nThe soft spots are real. The biggest one is the unspecified protocol for \\hat{r}_{r_u_i} and \\hat{m}_{p_u_j} in Eq. (6). The denoiser takes a noisy vector and a timestep; each neighbor has its own input. Eq. (8) samples only the query user's x_t, with no joint noise distribution for the neighbors, no statement about whether neighbor predictions are computed on the fly or cached, and no description of an inference-time approximation. Running the diffusion chain for every neighbor at every step would be far more expensive than the reported 10% overhead, so something is missing. Second, pseudo-user vectors are continuous TF-IDF values in [0,1] while the denoiser is trained on binary user vectors; there is no explicit pseudo-user reconstruction loss. The model may learn to handle this implicitly, but that is not explained. Minor issues: the text says top-20 neighbors are best but Table 4 shows Yelp preferring top-50; alpha, beta, gamma values are not reported per dataset; and the paired t-tests lack procedural detail (number of runs, error bars). None of these are fatal by themselves, but together they prevent reproduction.\n\nThis is a five-page workshop-style paper with a nice idea and plausible, small effect sizes. It does not deserve rejection on the merits, but it does need a precise statement of the training/inference algorithm before the numbers can be accepted. Send it to serious review with a request for that specification and for the per-dataset hyperparameters.","headline":"A nice pseudo-user + neighbor blend for diffusion recommenders, but the math as written doesn't define how neighbor predictions are computed; the consistent gains are therefore hard to trust without a reproducibility pass.","tokens_in":9801,"tokens_out":4457,"would_cite":false,"duration_ms":38282,"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":"Neighbor signals folded into diffusion denoising lift top-N recommendation accuracy.","keywords":["diffusion model","recommender system","collaborative filtering","item-side information","pseudo-user","neighborhood-based recommendation","implicit feedback","top-N recommendation"],"falsifier":"Replace each neighbor's denoised prediction in Eq. (6) with that neighbor's observed (noiseless) interaction vector, keeping the same neighbor sets and blend weights; if Recall@20 does not drop to the level of the single-source ablations, then the improvement does not actually come from denoising the neighbors.","tokens_in":8588,"feed_emoji":"🎯","tokens_out":5363,"duration_ms":47283,"temperature":0.7,"pith_summary":"Diffusion-based recommender systems typically train a denoiser to reconstruct a user's binary interaction history from Gaussian noise, but heavy noise can wash out the personal detail that matters for recommendation. This paper claims that the lost personalized information can be recovered by blending the user's own denoised vector with the denoised vectors of two kinds of neighbors: real users with similar behavior, and 'pseudo-users' created from item review words. The paper introduces CDiff4Rec, which builds these neighbor sets by cosine similarity and incorporates their preference information directly into the diffusion reconstruction objective. On three public datasets (Yelp, Amazon-Game, Citeulike-t), the full model beats six baselines on Recall@20 and NDCG@20, with the largest gains over the plain diffusion recommender that it extends. If this holds, it means item content and collaborative signals can be folded into the diffusion denoising objective without changing the backbone architecture.","feed_headline":"Diffusion recommender lifts accuracy by blending in neighbor signals","feed_subtitle":"Review-word pseudo-users and similar neighbors added to the denoising target beat all baselines on three public datasets.","key_machinery":"The machinery is a modified diffusion reconstruction target that averages the denoiser's output for the query user with weighted outputs for behavior-similar real neighbors and review-word pseudo-neighbors. Pseudo-users are generated by treating each review word as a user whose interaction vector is the TF-IDF weighted, min-max normalized word co-occurrence across items, making the vector continuous in $[0,1]^{|I|}$. Top-$K$ neighbor sets for every user are precomputed with cosine distance and cached, so the extra cost during training and inference is just the aggregation itself. The same denoiser MLP that reconstructs the query user also produces the neighbor predictions, and the blended prediction replaces the plain reconstruction in the diffusion loss, which is what forces the model to encode collaborative and content signals.","core_discovery":"CDiff4Rec's central claim is that the diffusion denoising objective should not stop at the user's own corrupted vector. Instead, the model defines a fine-grained reconstruction $\\hat{\\mathbf{r}}'_u$ as a weighted blend of the user's own denoiser output, the denoiser outputs of the top-$K$ most behaviorally similar real users, and the denoiser outputs of top-$K$ pseudo-users (review words with normalized TF-IDF interaction vectors). The weights $\\alpha$, $\\beta$, and $\\gamma$ sum to one, and attention scores among neighbors can be average pooling, precomputed cosine similarity, or learned parametric attention. The training loss in Eq. (8) is the same diffusion loss as the base DR, but with $\\hat{\\mathbf{r}}'_u$ in place of the plain single-user reconstruction, so the denoiser is trained to make the blended prediction match the user's true interaction vector. The experiments show this single change improves Recall@20 and NDCG@20 over all baselines on all three datasets, and that both neighbor types contribute, with the largest gains when combined.","pith_inferences":["The pseudo-user recipe should generalize to any item-side feature source (tags, categories, visual attributes), since it only needs a feature-item co-occurrence matrix and a normalization step.","Because neighbor sets are precomputed, this collaborative blend could be grafted onto future denoising backbones without retraining the neighbor selection, making it a plug-in training objective.","A direct test of whether the benefit comes from the blended target or from the shared denoiser would be to compute neighbor predictions with a separate, frozen denoiser while training the query denoiser; the paper leaves this compartmentalization unspecified.","The reported sensitivity to $K$ and the number of pseudo-users suggests an automatic selection rule based on validation recall could squeeze out further gains, but that is not explored in the paper."],"forward_implications":["CDiff4Rec improves top-20 recall and NDCG over the strongest baselines on Yelp, AM-Game, and Citeulike-t, and both neighbor types contribute, with the largest gains when real and pseudo neighbors are combined.","The accuracy gain over DiffRec is consistent across R@10, R@50, and R@100 with modest wall-clock overhead (about 10% on Yelp and near-zero on AM-Game), so the collaborative blend is cheap to add.","Strong performance is reached with only 1,000 pseudo-users and top-20 neighbors, meaning the precomputed neighbor sets keep the method scalable.","Because pseudo-users come from item-side features, item content can be exploited without adding a new loss term or a separate network for content.","The choice among average, behavior-similarity, and parametric attention lets practitioners trade simplicity, efficiency, and effectiveness, with behavior-similarity attention giving the best reported results."],"supporting_citations":[{"why":"Supplies the base diffusion recommender formulation (forward, backward, training, inference) that CDiff4Rec extends by adding neighbor preference information.","marker":"[20]"},{"why":"Provides the denoising diffusion probabilistic model foundations (noise schedule, ELBO, parameterized backward process) that the DR formulation relies on.","marker":"[8]"},{"why":"The strongest non-diffusion baseline in the comparisons; its neighborhood-based generative approach is the direct motivation for injecting collaborative signals.","marker":"[18]"},{"why":"A standard graph-based collaborative filtering baseline that represents the traditional CF performance CDiff4Rec aims to surpass.","marker":"[6]"},{"why":"A generative VAE-based CF baseline; CDiff4Rec's improvement over it shows the effect of combining content and collaborative signals under a diffusion objective.","marker":"[12]"},{"why":"A baseline that also uses item features through a conditioned variational autoencoder, providing a direct comparison for the item-side information component.","marker":"[1]"},{"why":"A classic implicit-feedback matrix factorization baseline; its ranking loss and performance serve as a lower-bound reference for top-N recommendation quality.","marker":"[16]"}],"fun_headline_variants":["Neighbor-blended diffusion target boosts recommender accuracy","CDiff4Rec: diffusion with real and pseudo neighbors tops baselines","Blending neighbor signals into diffusion denoising improves top-K recall","Collaborative diffusion model beats baselines by adding pseudo-users"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a denoiser trained on binary user interaction vectors also produces accurate denoised predictions for real neighbors and for continuous TF-IDF pseudo-user vectors, since the final reconstruction blends those predictions and is trained to match the user's original vector.","fun_headline_variants_meta":{"raw":{"variants":["Neighbor-blended diffusion target boosts recommender accuracy","CDiff4Rec: diffusion with real and pseudo neighbors tops baselines","Blending neighbor signals into diffusion denoising improves top-K recall","Collaborative diffusion model beats baselines by adding pseudo-users"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":2076,"prompt_tokens":897,"completion_tokens":1179,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":1108}},"tokens_in":513,"tokens_out":1179,"duration_ms":9721,"temperature":1.0,"reasoning_tokens":1108,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T21:41:21.274191+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Replace each neighbor's denoised prediction in Eq. (6) with that neighbor's observed (noiseless) interaction vector, keeping the same neighbor sets and blend weights; if Recall@20 does not drop to the level of the single-source ablations, then the improvement does not actually come from denoising the neighbors.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A baseline that also uses item features through a conditioned variational autoencoder, providing a direct comparison for the item-side information component."}],"review_version":1}