{"id":"81f37d4c-38b8-43b1-ad8b-1ea4ab0bdd45","arxiv_id":"2411.09181","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DeBaTeR adds timestamps into graph recommender embeddings and uses the augmented scores to reweight edges or losses, reporting modest accuracy and robustness gains over existing denoising methods.","lead":"Researchers add learned timestamp embeddings to user and item representations so graph-based recommenders can downweight clicks that look temporally out of place. On four public datasets, the method posts small gains over strong baselines, mostly when the injected noise is temporally inconsistent with an item's usual popularity pattern.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The additive time-aware score in Eq. (22) cannot represent the paper's motivating user-side temporal patterns: for fixed user and timestamp, e_u^T e_t is constant across items, so only global item-time popularity survives in ranking; a synthetic per-user time-flip test would settle it.","rationale":"The reader's weakest assumption is the same as the most load-bearing concern I find: the additive timestamp embedding in Eq. (15) cannot realize the paper's stated user-side temporal preference mechanism. The algebra is decisive. For fixed u and t, e_u^T e_t is a scalar that does not depend on the item, so it cannot affect item ranking; the only time-dependent ranking signal is e_i^T e_t, which is global item-level temporal popularity. The paper's introduction example ('horror at night, talk shows in the morning') requires a user x item x time interaction, which the decomposition in Eq. (14) cannot express. This is a correctness risk for the central interpretative claim, not merely a disagreement with current consensus. I also credit the paper where it has independent support: the noise-sampling experiments in Table 8 are a useful, falsifiable probe, and they show the method can detect item temporal incoherence, which is consistent with the surviving e_i^T e_t term. That partial positive result does not fix the user-side gap. The ablation results in Tables 5 and 6 are mixed rather than uniformly supportive, especially for DeBaTeR-L on ML-1M, so the empirical case is weaker than the prose suggests. Nevertheless, the framework could plausibly be repositioned as an item-side temporal denoising method, and the paper's main utility results might survive after re-scoping. Thus the existing CONDITIONAL verdict is appropriate; I would not move it to accept or reject without the proposed synthetic test and a revised statement of what the model can and cannot capture.","tokens_in":25167,"tokens_out":3442,"duration_ms":47365,"concrete_test":"Build a synthetic dataset with, e.g., 200 users, two item genres, and two time slots (morning/night), where each user's genre preference flips by slot and item-level global time popularity is balanced across genres. Train DeBaTeR-A and DeBaTeR-L with the same hyperparameter settings as in Section 4.1 and evaluate whether morning top-k rankings favor genre A and night rankings favor genre B. If per-user time-flip accuracy stays near chance, the additive score cannot encode user-side temporal patterns, confirming the concern. If accuracy is consistently high, the additive form must be capturing the pattern through some indirect mechanism, and this attack should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.1 claims that Eq. (14) is capable of capturing both user-side and item-side temporal patterns, and the introduction's central motivating example is user-specific ('this user watches horror at night and talk shows in the morning'). But the actual ranking score in Eq. (22) is P_{u,i}^t = (e_u+e_t)^T(e_i+e_t) = e_u^T e_i + e_u^T e_t + e_i^T e_t + ||e_t||^2. For a fixed user u and timestamp t, the term e_u^T e_t and the norm term are constant across items i, so they cancel in any item-vs-item comparison: P_{u,i}^t - P_{u,j}^t = e_u^T(e_i-e_j) + e_t^T(e_i-e_j). There is no third-order user-item-time interaction. The model therefore expresses only static user preference plus global item-time popularity; it cannot express a per-user preference that changes with time of day, because such a signal requires the effect of t to differ across users. This is not a tuning or approximation issue: it is structurally absent from Eq. (15). The RQ3 experiments support item-side temporal coherence detection, which is exactly what e_i^T e_t can provide, but they do not rescue the user-side claim. The ablation tables add further ambiguity: on ML-1M vanilla and noisy data, removing time from losses and prediction sometimes improves Precision/Recall for DeBaTeR-L, so even the utility claim is not uniformly supported. The safest conclusion is that the framework may denoise via item temporal coherence, while the paper's headline user-side temporal mechanism is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DeBaTeR, a mechanism that adds a learned timestamp embedding to user and item embeddings in neural graph collaborative filtering, and uses the resulting time-aware embeddings in two denoising strategies: reweighting the adjacency matrix (DeBaTeR-A) and reweighting the loss function (DeBaTeR-L). The authors argue that temporal incoherence is a useful signal for identifying noisy interactions, motivated by examples such as a user who watches horror movies at night and talk shows in the morning. Experiments on ML-100K, ML-1M, Yelp, and Amazon compare the proposed methods against general and denoising recommender baselines, including an ablation study and a noise-sampling analysis (RQ3).","tokens_in":25568,"tokens_out":6770,"duration_ms":67082,"significance":"If the proposed mechanism works as claimed, it is a simple, general way to incorporate timestamps into graph collaborative filtering and could improve both recommendation utility and robustness to noisy implicit feedback. The paper's RQ3 experiments provide a falsifiable, external check that item-side temporal coherence is being used for denoising, and the authors state that code is provided. However, the headline user-side temporal-pattern claim is not realized by the model, and the ablation results show that time-aware embeddings do not consistently improve utility. The item-side temporal denoising insight is a useful contribution, but the manuscript needs substantial reframing and additional validation before the broader claims are supported.","major_comments":[{"comment":"The claim that the time-aware preference in Eq. (22) captures both user-side and item-side temporal patterns is not supported by the model's functional form. For a fixed user u and timestamp t, the score difference between items i and j is P_{u,i}^t - P_{u,j}^t = (e_u + e_t)^T(e_i - e_j), because e_u^T e_t and ||e_t||^2 are constant across items. Thus the model cannot represent a per-user temporal preference that changes the ranking of items differently for different users; it can only express a global item-time popularity term e_t^T e_i. The introduction's motivating example (a user watches horror at night and talk shows in the morning, so a morning horror viewing is noisy) requires a user-time interaction that is structurally absent from Eq. (15). This is not a tuning issue: the additive architecture cannot realize the stated motivating scenario. The RQ3 experiments in Section 4.4 support item-side temporal coherence, but they do not rescue the user-side claim.","section":"Section 3.1, Eqs. (14)-(15) and (22)"},{"comment":"The ablation narrative is contradicted by the ML-1M results for DeBaTeR-L. On the vanilla ML-1M set (Table 5), removing time from losses and prediction improves Precision@10 from 0.1501 to 0.1667 and Recall@10 from 0.0404 to 0.0477; on the noisy ML-1M set (Table 6), Precision@10 improves from 0.1538 to 0.1682 and Recall@10 from 0.0422 to 0.0504. The text in Section 4.3 states that removing time-aware embeddings from loss and prediction leads to an overall worse performance on vanilla datasets, which is not true for this dataset. The paper should either qualify this claim or provide a metric-level explanation (e.g., NDCG is higher with time, while precision/recall are not). As written, the utility benefit of time in DeBaTeR-L is inconsistent and weakens the abstract's assertion that time information 'indeed helps'.","section":"Section 4.3, Tables 5 and 6"},{"comment":"Several of the claimed improvements over the BOD baseline are within one standard deviation, and the paper does not report significance tests. For example, in Table 2 on ML-1M, NDCG@10 is 0.1903 (std 0.0028) for DeBaTeR-A versus 0.1901 (std 0.0025) for BOD, and in Table 3 on ML-1M, NDCG@10 is 0.1865 (std 0.0074) for DeBaTeR-L versus 0.1849 (std 0.0052) for BOD. Similar overlaps occur on Yelp and Amazon for several metrics. The paper's statement that the methods 'outperform state-of-the-art models for both utility and robustness' is therefore not uniformly supported by the reported statistics; the authors should either add significance testing, report effect sizes, or temper the claim to specifically identified metrics and datasets.","section":"Tables 2 and 3, Section 4.2"}],"minor_comments":[{"comment":"Please specify the exact baseline set and averaging procedure for the relative improvement percentages in Table 7, since it is currently unclear whether the percentages are averaged over all datasets, all metrics, or all baselines.","section":"Section 4.2, Table 7"},{"comment":"The reliability score uses zero-layer embeddings e_u^(0) and e_i^(0), while the cited method [29] uses first- and second-layer embeddings; please justify why the unpropagated initial embeddings are sufficient for detecting noisy edges.","section":"Section 3.2, Eq. (19)"},{"comment":"Please clarify how the negative-sample weight w_{u,j} is generated; the notation in Eq. (28) suggests W(e_u || e_i || e_t), but Eq. (29) uses w_{u,j} without specifying whether the same function is applied with the negative item embedding in the second argument.","section":"Section 3.3, Eqs. (28)-(29)"},{"comment":"Algorithm 2 line 6 derives embeddings from Eq. (2), whereas DeBaTeR-A uses the time-aware propagation in Eq. (20); please clarify whether DeBaTeR-L's backbone message passing also uses time-aware embeddings or only the losses and prediction.","section":"Algorithm 2, line 6"},{"comment":"The statement that this is 'the first work to leverage time information in neural graph collaborative filtering recommender systems' is too broad given the existence of temporal graph recommender methods such as [1]; please restrict the novelty claim to the denoising setting or to the specific mechanism proposed.","section":"Introduction, Section 1"},{"comment":"The subscripts in Eq. (26) and (27) are difficult to parse (e.g., u_i_t_ui); please use a consistent notation such as t_{ui} for the timestamp of the positive interaction to distinguish it from generic timestamps.","section":"Section 3.2, Eq. (26)"}],"recommendation":"major_revision","confidential_remarks":"The paper's most defensible contribution is item-side temporal denoising, and the RQ3 noise-sampling experiments provide useful external validation. The user-side temporal-pattern story in the introduction and Section 3.1 is not realized by the additive model, and the ablation inconsistencies on ML-1M undermine the utility claim as currently worded. I would encourage the authors to reframe the contribution around item-side temporal coherence, consider adding a per-user temporal interaction term if they wish to keep the user-side narrative, and be more careful about significance claims. The manuscript is not beyond repair, but the central claims need substantial revision before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What's actually new here is a modest, useful result: temporal incoherence is a detectable noise signal, and you can inject timestamps into existing denoising frameworks (structure denoising [29] and BOD's loss reweighting [26]) with a simple additive encoder. The RQ3 experiment, which constructs noisy datasets by sampling items against hourly popularity, is the best part of the paper—it tells you what the method actually detects. Don't let the title's 'user-side temporal patterns' fool you: the model's ranking score can't represent them.\n\nThe math is straightforward. Eq. (22) gives P^t_{u,i} = (e_u+e_t)^T(e_i+e_t) = e_u^T e_i + e_u^T e_t + e_i^T e_t + ||e_t||^2. For a fixed user u and time t, the e_u^T e_t term is constant across items, so ranking is e_u^T e_i + e_i^T e_t plus a constant. That is static preference plus global item-time popularity. The horror-at-night/talk-show-in-morning example in the introduction requires a user-specific time interaction, which is structurally absent. Section 3.1's claim that Eq. (14) captures both user-side and item-side temporal patterns is therefore an overstatement. The RQ3 experiments only probe item-side patterns, and they support that reading.\n\nOther soft spots are proportionate. The ML-1M ablation is genuinely mixed: DeBaTeR-L with time removed from losses and prediction gets better Precision@10 (0.1667 vs 0.1501) and Recall@10 (0.0477 vs 0.0404), though NDCG@10 is slightly worse. So 'time always helps' is not established. Gains over BOD are often within one standard deviation on several metrics, so the SOTA claim should be softened. The code is promised but not linked in the arXiv version.\n\nWhat the paper does well: it's clearly written, the setup is fair (multiple datasets, standard baselines, official repos for the denoising baselines), and the timestamp encoder is cheap to add to any dot-product CF model. The RQ3 methodology—varying noise distributions to see which ones are detected—should be copied by other denoising papers.\n\nBottom line: this is an incremental but real contribution to item-side temporal denoising, and the headline user-side mechanism is unsupported. A synthetic test (flipping a user's item types across time slots) would settle the question. I'd send it to peer review: the item-side finding is worth publishing and the user-side claim is fixable by rewriting the motivation and claims. I'd want a major revision before acceptance.","headline":"Useful item-side temporal denoising result, but the paper's user-side temporal pattern story is algebraically unsupported and needs to be cut or fixed.","tokens_in":26104,"tokens_out":4442,"would_cite":true,"duration_ms":42375,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding a learned timestamp vector to user and item embeddings lets graph collaborative filtering detect noisy interactions, improving utility and robustness.","keywords":["denoising recommender systems","graph collaborative filtering","time-aware embeddings","temporal patterns","implicit feedback","bipartite graph","robustness","noise injection"],"falsifier":"Take a dataset where item popularity is uniform across time but each user has strong, known time-of-day preferences, and compare DeBaTeR against the same backbone without timestamps; if the claimed temporal mechanism is working, ranking accuracy should improve, and if the additive user term is inert, it will not. A second check is to inject noise that follows user-side temporal incoherence (a user's favorite item category moved to the wrong time of day) rather than item-popularity noise; the claimed advantage of DeBaTeR should shrink in that setting.","tokens_in":24942,"feed_emoji":"🕒","tokens_out":7864,"duration_ms":74103,"temperature":0.7,"pith_summary":"This paper tries to establish that interaction timestamps are a usable signal for cleaning noisy implicit-feedback data in graph collaborative filtering, where clicks or purchases often do not reflect real preference. It proposes DeBaTeR, a mechanism that forms time-aware user and item embeddings by adding the same learned timestamp embedding to each, and then uses the dot product of these time-aware embeddings as a reliability score in two denoising strategies: reweighting the adjacency matrix (DeBaTeR-A) and reweighting the loss function (DeBaTeR-L). The paper's claim is that temporal incoherence, an interaction happening at a time when that item's recorded activity is low, marks noisy edges, and that accounting for this improves accuracy and robustness to injected noise. If true, timestamps become a cheap, already-collected signal for denoising, and the additive embedding trick can be dropped into existing neural graph collaborative filtering models.","feed_headline":"Timestamps help recommender systems spot noisy clicks","feed_subtitle":"Adding a timestamp vector to user and item embeddings improves accuracy and robustness on four real datasets.","key_machinery":"The central object is the additive time-aware embedding pair, $e_u^t = e_u + e_t$ and $e_i^t = e_i + e_t$, where $e_t$ is learned by one-hot encoding each timestamp dimension and concatenating per-dimension embeddings (Eqs. 16-18). The corresponding score $P_{u,i}^t = e_u^\\top e_i + e_u^\\top e_t + e_i^\\top e_t + \\|e_t\\|^2$ is the time-aware preference used in prediction. The same construction is plugged into a cosine-based reliability score for adjacency-matrix reweighting in DeBaTeR-A and into an MLP weight generator trained by gradient matching in DeBaTeR-L, so the timestamp encoder is the piece that carries the temporal signal through both denoising strategies.","core_discovery":"DeBaTeR claims that a timestamp embedding $e_t$, added to both sides of the interaction as $e_u^t = e_u + e_t$ and $e_i^t = e_i + e_t$, turns the prediction score $P_{u,i}^t = (e_u + e_t)^\\top (e_i + e_t)$ into a sum of a general preference term and a temporal term. The paper argues that this decomposition lets the model capture temporal patterns of users and items, and it uses the same time-aware embeddings in a cosine reliability score for pruning or downweighting graph edges (DeBaTeR-A) and in an MLP weight generator for reweighting training samples (DeBaTeR-L). On four real-world datasets, one of the two variants reaches the best precision, recall, or NDCG in most settings, beating state-of-the-art graph collaborative filtering and denoising baselines on both vanilla data and data with 20 percent injected noise. The paper presents this as the first use of time information for denoising in neural graph collaborative filtering.","pith_inferences":["A direct reading of the score formula suggests the reported gains are most plausibly carried by item-side temporal popularity: for a fixed user and time, the term $e_u^\\top e_t$ is constant across items, so user-specific patterns such as 'this user watches horror at night' cannot change the item ranking.","A natural test of the mechanism is to construct a synthetic dataset with strong user-side temporal preferences but flat item-side popularity; if the additive model is what the paper says it is, it should still show an advantage, and if only item-side terms matter, it will match a timestamp-free baseline.","The discrete per-dimension timestamp encoder could be replaced by continuous time embeddings or by separating user-time and item-time vectors, which would let the model actually express user-specific temporal preferences rather than only item popularity by time.","Temporal reliability is an orthogonal signal to loss-based and agreement-based denoising, so combining DeBaTeR's reliability scores with those families could yield further robustness gains, though the paper does not test such combinations."],"forward_implications":["If the central claim is correct, recommender systems can treat an interaction's timestamp as evidence about whether the interaction is noisy, without needing any explicit feedback label.","The additive embedding trick is architecture-light: existing graph collaborative filtering models can adopt it by adding the timestamp vector to user and item embeddings before scoring and loss computation.","Temporal coherence becomes a usable pruning signal for the bipartite interaction graph, complementing existing denoising methods that rely on embedding similarity or loss magnitude.","Item-side periodic popularity, such as seasonal items clustering at particular times of year, becomes directly learnable from timestamps and can improve ranking at query time.","The method's robustness gains on data with injected random noise suggest it can partially absorb poisoning attacks that place interactions at arbitrary timestamps."],"supporting_citations":[{"why":"Supplies the graph collaborative filtering backbone, message-passing equations, and contrastive learning objectives that DeBaTeR builds on.","marker":"[30]"},{"why":"Provides the reliability-score reweighting of the adjacency matrix that DeBaTeR-A generalizes with time-aware embeddings.","marker":"[29]"},{"why":"Supplies the weight-generator and gradient-matching training scheme that DeBaTeR-L extends with timestamp embeddings.","marker":"[26]"},{"why":"Provides the per-dimension one-hot timestamp embedding construction that DeBaTeR adapts for its timestamp encoder.","marker":"[17]"},{"why":"Defines the BPR ranking loss that DeBaTeR reweights in both denoising strategies.","marker":"[18]"},{"why":"Provides the alignment-and-uniformity (AU) loss used in DeBaTeR's combined objective and in the gradient-matching objective.","marker":"[21]"},{"why":"Provides two of the timestamped movie-rating datasets used for evaluation.","marker":"[6]"},{"why":"Provides the timestamped product review dataset used for evaluation on a larger sparse interaction graph.","marker":"[15]"}],"fun_headline_variants":["Adding timestamps to embeddings reveals noisy edges in recommenders","DeBaTeR: denoising graph recommenders with time-aware weights","Timestamps decompose predictions to expose noisy interactions","A temporal term helps recommenders spot and fix noisy clicks","Time information in embeddings helps prune noisy graph edges"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that adding the same learned timestamp vector to both user and item embeddings lets the model exploit temporal patterns on either side, but because $e_u^\\top e_t$ is constant across items for a fixed user and timestamp, user-specific patterns like 'horror at night, talk shows in the morning' cannot affect the ranking of items, leaving only item-side temporal popularity to carry the signal.","fun_headline_variants_meta":{"raw":{"variants":["Adding timestamps to embeddings reveals noisy edges in recommenders","DeBaTeR: denoising graph recommenders with time-aware weights","Timestamps decompose predictions to expose noisy interactions","A temporal term helps recommenders spot and fix noisy clicks","Time information in embeddings helps prune noisy graph edges"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000956,"raw_usage":{"total_tokens":4120,"prompt_tokens":1032,"completion_tokens":3088,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":648,"completion_tokens_details":{"reasoning_tokens":3008}},"tokens_in":648,"tokens_out":3088,"duration_ms":47350,"temperature":1.0,"reasoning_tokens":3008,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:57:48.085261+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a dataset where item popularity is uniform across time but each user has strong, known time-of-day preferences, and compare DeBaTeR against the same backbone without timestamps; if the claimed temporal mechanism is working, ranking accuracy should improve, and if the additive user term is inert, it will not. A second check is to inject noise that follows user-side temporal incoherence (a user's favorite item category moved to the wrong time of day) rather than item-popularity noise; the claimed advantage of DeBaTeR should shrink in that setting.","supporting_citations":[{"cited_title":"Are graph augmentations necessary? simple graph co ntrastive learn- ing for recommendation","cited_arxiv_id":null,"evidence_quote":"Supplies the graph collaborative filtering backbone, message-passing equations, and contrastive learning objectives that DeBaTeR builds on."},{"cited_title":"Toward s robust neural graph collaborative filtering via structure denoising and e mbedding perturba- tion","cited_arxiv_id":null,"evidence_quote":"Provides the reliability-score reweighting of the adjacency matrix that DeBaTeR-A generalizes with time-aware embeddings."},{"cited_title":"Efficient bi-level optimization for recommendation denoisi ng","cited_arxiv_id":null,"evidence_quote":"Supplies the weight-generator and gradient-matching training scheme that DeBaTeR-L extends with timestamp embeddings."},{"cited_title":"Incorpo rating time in se- quential recommendation models","cited_arxiv_id":null,"evidence_quote":"Provides the per-dimension one-hot timestamp embedding construction that DeBaTeR adapts for its timestamp encoder."},{"cited_title":"Bpr: Bayesian personalized ranking from implicit f eedback","cited_arxiv_id":null,"evidence_quote":"Defines the BPR ranking loss that DeBaTeR reweights in both denoising strategies."},{"cited_title":"Maxwell Harper and Joseph A","cited_arxiv_id":null,"evidence_quote":"Provides two of the timestamped movie-rating datasets used for evaluation."},{"cited_title":"Justifying recommendations using distantly-labeled reviews and fine-grained aspects","cited_arxiv_id":null,"evidence_quote":"Provides the timestamped product review dataset used for evaluation on a larger sparse interaction graph."}],"review_version":1}