{"id":"3cfe3d04-197a-486c-99c8-85d0ee5d5457","arxiv_id":"2412.11589","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"FENRec adds time-decayed future-item soft labels and anchor-mixed enduring hard negatives to contrastive sequential recommenders, improving next-item prediction on four datasets.","lead":"This paper proposes FENRec, a training technique for sequential recommendation that uses a user's near-future interactions as soft labels and mixes anchor sequences with random negatives to create persistently harder negative examples. If the reported results hold, it gives a simple plug-in improvement of a few percent on standard benchmarks for contrastive recommenders.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Potential label leakage from time-dependent soft labels: Eq. (6) reads future positions from the full user sequence, so under the standard hold-last-two split the subsequence ending at T-2 is trained on the validation and test items, which alone could explain the 6.16% SOTA gain.","rationale":"FENRec's central claim is empirical SOTA; the first contribution is soft labels from future interactions. For that claim to be true, the future interactions used as training labels must be drawn from the training split, not from the portions of the user sequence held out for validation and test. The paper's notation does not guarantee this. Eq. (6) uses S_u, introduced as the full interaction sequence, and no masking is described. Given the near-universal leave-one/two-out protocol for Amazon and Yelp benchmarks, the maximum training subsequence that the authors would naturally generate ends at T-2 (so its immediate next item T-1 is the validation item or, in some pipelines, the test item T itself); with two extra future positions, both held-out items are inside the soft-label window. The reported gains, especially the 6.34%/5.99% over the best baseline and the clean monotone improvements in Table 3, are exactly what would be expected if the model is trained to put probability mass on the evaluation labels. The reader's baseline-tuning concern is legitimate but secondary: even if all baselines were perfectly tuned, label leakage would invalidate the comparison, whereas fixing the leakage would not require re-tuning baselines. I therefore disagree with the reader's choice of weakest assumption. I am not asserting the leak is present; I am asserting the paper as written does not rule it out, and the corrupted code link makes the omission uncheckable. The geometric lemma and the hard-negative construction are internally plausible, and the compatibility results in Table 2 would be meaningful if the leakage check passes, so the paper should be conditionally accepted pending this diagnostic.","tokens_in":21431,"tokens_out":15118,"duration_ms":144211,"concrete_test":"Fix the code link and run a single diagnostic: for every training subsequence used in FENRec on the four datasets, print the set of item positions that receive nonzero soft-label mass in Eq. (6) and check whether any position > |S_u|-2 (the held-out validation/test region under the standard split) appears. Then retrain FENRec with the soft-label support clamped to positions ≤ |S_u|-2, keeping all hyperparameters identical. If the clamped model retains the Table 1 improvements, the concern is resolved; if performance drops to roughly the FENRec-S ablation level or to baseline, the SOTA claim is an artifact of label leakage.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim is the 6.16% average improvement over baselines; the load-bearing premise is that time-dependent soft labels in Eq. (6) do not use items held out for validation/test. The paper defines S_u in §Preliminaries as the full chronological user interaction sequence, and Eq. (6) assigns soft-label mass to positions |S_us|+1 through min(|S_us|+3, |S_u|+1). No sentence in the main text or Appendix limits these 'future' positions to the training split. The baselines are run with the standard leave-last-item-for-test / second-to-last-for-validation protocol (the Appendix cites SSLRec and author code for most baselines). Under that protocol, the training subsequence ending at position T-2 has a soft-label window {T-1, T}; the subsequence ending at T-3 has {T-2, T-1, T}. Those are exactly the validation and test items that FENRec is later required to rank. If the released implementation uses the full sequence as S_u, the model is trained to put probability mass on the labels it is then evaluated against, so the claimed 2-12% relative gains and the ablation results in Table 3 would not measure generalization. This is an internal gap: Implementation Details never specifies the masking of soft-label positions relative to the split, and the corrupted code link prevents checking. It is not a disagreement with community consensus; it is a question about whether the experiments test the stated hypothesis.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FENRec, an add-on for contrastive sequential recommendation. It replaces binary next-item labels with time-dependent soft labels built from the next two future interactions in the user's chronological sequence, and it augments the contrastive loss with enduring hard negatives obtained by mixing the anchor with in-batch negatives, using stop-gradient and a focal-style upweighting term. The method is evaluated by inserting it into ICSRec, CL4SRec, and DuoRec on Amazon Sports/Beauty/Toys and Yelp, and the paper reports an average relative improvement of 6.16% over baselines, together with ablations and compatibility experiments.","tokens_in":21737,"tokens_out":5814,"duration_ms":51604,"significance":"The two proposed ingredients are simple and potentially portable, and the geometric claim that a mixed negative has higher inner product with the anchor than the original negative is proved in Appendix Lemma 2 and is correct. The compatibility study in Table 2 is a genuine strength: it shows the components can be wrapped around existing contrastive SR frameworks rather than requiring a new backbone. However, the central empirical claim rests on a training/evaluation protocol that is not fully specified. As written, the soft-label construction in Eq. (6) appears to use held-out validation and test items as training labels, which would invalidate the reported state-of-the-art results and the ablation conclusions. The baseline tuning asymmetry further weakens the quantitative comparison, and the broken code link prevents independent verification of the implementation.","major_comments":[{"comment":"The soft-label construction is not described as being restricted to the training portion of each user sequence. Eq. (6) assigns probability mass to positions |S_us|+1 through min(|S_us|+3, |S_u|+1) of the full chronological sequence S_u, and the paper never states that these future positions are masked with respect to the train/validation/test split. Under the standard leave-last-two protocol used by the baselines, a subsequence ending at position T-2 receives soft-label mass on items T-1 and T, which are exactly the validation and test items on which FENRec is later ranked. The abstract's code link is corrupted (\"zurlthttps://github.com/uikdwnd/FENRecu\"), so the reader cannot check the released implementation. As written, the 6.16% average improvement and the Table 3 ablations do not demonstrate generalization; they may reflect training on the evaluation labels. Please specify and implement the masking of soft-label positions relative to the split and re-run all experiments, or provide explicit evidence that no held-out item ever appears in a soft-label window.","section":"Method, Eq. (6); Appendix, Implementation Details"},{"comment":"The claim that baselines are reported \"under its optimal settings\" is not supported by the described procedure. Baseline hyperparameters are taken from the original papers or public implementations, while FENRec's own hyperparameters (γ, λ, τ2, m, α, number of future items, and warm-up length) are tuned per dataset in the main text and in Appendix Figure G. Without a comparable per-dataset tuning protocol for the baselines, the reported relative gains (for example, 10-12% on Yelp) are not a fair state-of-the-art comparison, and the abstract's average improvement figure may be inflated. Please document the baseline tuning procedure or revise the optimality claim.","section":"Appendix, Implementation Details"},{"comment":"The main text states that FENRec \"outperforms all other models across all metrics,\" but the significance annotations in Table E show that several best results are not marked with an asterisk (e.g., Beauty HIT@20 and NDCG@20; Toys HIT@5 and NDCG@5). The Appendix itself says improvements are significant \"across most metrics.\" The abstract and main-text claim should be qualified to match the statistical evidence, unless the missing markers are an error.","section":"Table E and Comparison to SOTA (main text)"}],"minor_comments":[{"comment":"The code link is malformed: \"zurlthttps://github.com/uikdwnd/FENRecu\" should be \"https://github.com/uikdwnd/FENRec\".","section":"Abstract"},{"comment":"In case 1(b), the sentence \"θxz = θxy + π = 0\" is a typo; it should state that θxz = 0 and θxy = π.","section":"Appendix, Lemma 1 proof"},{"comment":"The first paragraph of Related Work refers to \"CL4Rec,\" while the rest of the paper uses \"CL4SRec\"; please unify the naming.","section":"Related Work"},{"comment":"\"Comparsion\" should be \"Comparison\".","section":"Figure 5 caption"},{"comment":"The phrase \"noise-based negative sampling method on sentence representations\" is vague; please specify how this method is adapted to sequential item representations and whether it is applied identically to all compared methods.","section":"Appendix, Implementation Details"},{"comment":"The ablation table reports only a single run configuration per component removal; adding significance tests or repeated-run variability would help support the claim that the performance drops are not due to noise.","section":"Table 3"}],"recommendation":"major_revision","confidential_remarks":"If the leakage concern is confirmed by the authors' code, the paper's central empirical claim would be an artifact and the manuscript should be rejected rather than revised. The current written protocol is consistent with a correctable experimental-design error, so I recommend major revision with a mandatory re-run of all experiments under a masked soft-label construction, plus a fair baseline tuning procedure. The broken code link and the missing significance on several metrics should also be addressed before resubmission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"FENRec is a competent empirical SR paper with a clever-looking trick, but a potential split-masking gap in the soft-label construction could undermine the headline 6.16% improvement. I would not accept the empirical claim as-is until that is checked.\n\nWhat is genuinely new: the paper combines two known ideas in a way that is new to SR contrastive learning — time-decayed soft labels drawn from the next two future interactions, and anchor-mixed enduring hard negatives with stop-gradient and focal upweighting. The geometric lemma (mixed negative has higher inner product with the anchor) is correct, and the paper is clearly written. Experiments on four standard benchmarks with multiple seeds and reported standard deviations are a cut above the field's typical practice.\n\nThe problem is load-bearing. Eq. (6) assigns soft-label mass to positions up to |S_u|+1, where S_u is defined as the full chronological user sequence. The paper never states that these future positions are restricted to the training split. Under the standard leave-last-two protocol, subsequences ending at T-2 and T-3 get soft-label mass on the validation and test items. If the implementation behaves as written, the model is trained to put probability mass on exactly the items it is later ranked against, and the ablation results would measure leakage rather than generalization. The code link in the abstract is corrupted (\"zurl\"), so the reader cannot check. This is the first thing a referee should demand.\n\nSecondary issues: baselines are reported at \"optimal settings\" but their hyperparameters are taken from the original papers while FENRec's are tuned per dataset; the noise-based negative sampling step is left unspecified; and not all metrics carry significance markers (e.g., Beauty HIT@20, NDCG@20 and several Toys metrics). These are real but minor compared to the leakage question.\n\nThe citation pattern looks fair, and the related-work coverage is adequate. The paper does not hide its components behind opaque tricks, and the math checks out on its own terms.\n\nWho this is for: people working on contrastive sequential recommendation who want a plug-in trick, and anyone interested in subtle train/validation leakage in sequence models. I would bring this to a reading group to discuss the methodological trap.\n\nRecommendation: send it to peer review. The idea deserves referee time, and the leakage question can be settled by a code inspection. If the soft labels are masked to the training interval, this is a solid incremental contribution; if not, the SOTA claim collapses.","headline":"FENRec is a plausible incremental trick for contrastive SR, but the paper never shows that its soft-label window is capped at the training split, so the headline 6.16% gain may be train/test leakage; referees should ask for the code before believing the empirical claim.","tokens_in":22322,"tokens_out":3027,"would_cite":false,"duration_ms":28893,"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":"This paper claims FENRec improves sequential recommendation by about 6.16% on average across four benchmark datasets by replacing binary targets with time-dependent soft labels drawn from future interactions and by injecting enduring hard…","keywords":["sequential recommendation","contrastive learning","soft labels","hard negative mining","data sparsity","future interaction utilization","top-k ranking"],"falsifier":"Re-run the main table giving all fourteen baselines the same per-dataset hyperparameter search budget as FENRec, with FENRec's own hyperparameters fixed before seeing test metrics; if the average improvement falls to the noise level, the central state-of-the-art claim is falsified.","tokens_in":21197,"feed_emoji":"🎯","tokens_out":7071,"duration_ms":60018,"temperature":0.7,"pith_summary":"The paper claims that data sparsity in sequential recommendation can be attacked more effectively by two cheap changes to contrastive training. Instead of rewarding only the exact next item, FENRec assigns decaying probabilities to the next two future interactions, turning upcoming behavior into soft labels. Instead of relying only on whatever in-batch negatives happen to appear, it manufactures enduring hard negatives by mixing each anchor with its negatives and reweighting the loss toward them. On Amazon Sports, Beauty, Toys, and Yelp, the combined method beats fourteen published baselines, with a claimed average improvement of 6.16% across Hit Ratio and NDCG at K=5, 10, and 20. A thoughtful reader would care because both components replace existing loss terms directly and can be dropped into contrastive sequential recommenders without new modules.","feed_headline":"Future data, harder negatives boost next-item hits 6 percent","feed_subtitle":"Two no-extra-module loss terms—soft future labels and enduring hard negatives—raise top-20 accuracy across four benchmarks.","key_machinery":"The two load-bearing mechanisms are Time-Dependent Soft Labeling (Eqs. 6-7), which replaces the one-hot next-item target with a geometric distribution over the next up-to-two future items, and Enduring Hard Negatives Incorporation (Eqs. 9-10), which constructs synthetic negatives from normalized anchor and in-batch negative representations as $\\tilde{h}_i = h_i/\\|h_i\\|_2$ and $h^-_{i,n} = \\lambda\\tilde{h}_i + (1-\\lambda)\\tilde{n}$, rescaled to the negative's norm. The paper proves a lemma that this construction keeps the synthetic negative closer in inner product to the anchor than the original negative, so it remains hard as training progresses. The final contrastive loss (Eq. 14) upweights these samples with a $\\tanh$ similarity gate controlled by $\\tau_2$ and a margin $m$, while `stop gradient` prevents the anchor from being pushed away from a vector that contains itself. Because these losses replace the standard $\\mathcal{L}_{rec}$ and $\\mathcal{L}_{cl}$, FENRec can be bolted onto any contrastive sequential recommendation framework.","core_discovery":"FENRec establishes that future interactions, normally discarded at training time, can be converted into time-dependent soft labels: for a subsequence ending at position t, the items at positions t+1 and t+2 receive probabilities that decay geometrically with a smoothing hyperparameter gamma. Coupled with this, the method generates enduring hard negatives by normalizing anchor and negative representations, mixing them with a weight lambda, renormalizing, and rescaling to the negative's norm; a lemma shows the mixed vector has a strictly higher dot product with the anchor than the original negative does, so the negatives stay challenging late in training. The contrastive loss then upweights these hard negatives through a tanh-based similarity reweighting with a margin, and a stop-gradient prevents the anchor from repelling a vector that contains itself. The paper reports that this recipe beats all baselines on all four datasets across every metric, and that removing either component degrades performance, with the soft-label term having the larger effect.","pith_inferences":["Because the soft-label term adds no extra module and depends only on the sequence itself, the same future-labeling trick should transfer to non-contrastive sequence encoders; if that holds, FENRec's contribution is not limited to contrastive frameworks.","The geometric decay controlled by gamma is effectively a label-smoothing schedule over future positions, which predicts that the optimal gamma should shift with dataset sparsity and average sequence length; this is testable across additional datasets.","The enduring-hard-negative guarantee is only about cosine closeness of a single mixed vector, and in-batch mixing can amplify false negatives as batch size grows, so the reported optimal batch size of 256 may not transfer to datasets with strong item co-occurrence.","The paper's sensitivity analysis shows lambda matters little for final performance, which suggests most of the observed gain comes from the time-dependent soft-label head rather than from the specific form of the hard-negative mixing."],"forward_implications":["Dropping FENRec's two loss terms into an existing contrastive sequential recommender, demonstrated on ICSRec, CL4SRec, and DuoRec, improves next-item ranking by roughly 1.7% to 12.3% relative depending on dataset and metric.","The gains are larger at K=20 than at K=5, so the method improves ranking quality deeper into the recommendation list, not just at the very top.","Users with very few interactions also benefit: FENRec outperforms every baseline in every interaction-length group, including cold-start users, on all four datasets.","The ablation study implies the soft-label component contributes more than the hard-negative component, and the two are synergistic because soft labels mitigate the false-negative risk that hard negatives introduce."],"supporting_citations":[{"why":"Supplies the mixing-of-hard-negatives strategy that FENRec adapts from image contrastive learning to user representations.","marker":"(Kalantidis et al. 2020)"},{"why":"Provides the mixed-negatives construction (MixCSE) that inspires the enduring hard negatives.","marker":"(Zhang et al. 2022)"},{"why":"Gives the focal InfoNCE reweighting that FENRec imports as its hard-negative upweighting contrastive loss.","marker":"(Hou and Li 2023)"},{"why":"DuoRec is the framework used in the compatibility experiments and the source of the uniformity metric.","marker":"(Qiu et al. 2022)"},{"why":"ICSRec is the backbone system into which FENRec is integrated for the main comparison table.","marker":"(Qin et al. 2024)"},{"why":"CL4SRec serves both as a representative contrastive baseline and as a second framework for the +FENRec compatibility test.","marker":"(Xie et al. 2022)"},{"why":"ICLRec sets the intent-contrastive baseline and supplies the dataset preprocessing conventions FENRec follows.","marker":"(Chen et al. 2022)"},{"why":"MVS is the label-smoothness baseline whose soft-label approach FENRec contrasts with its own future-based soft labels.","marker":"(Zhou et al. 2023)"}],"fun_headline_variants":["Soft future labels + enduring negatives lift SR by 6%","FENRec: future soft labels, hard negatives, +6% hits","Enduring negatives + future soft labels: +6% next-item","Future data as soft labels, durable negatives: +6% accuracy","Soft future labels + hard negatives: 6% better next-item"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed 6.16% average improvement presupposes a fair comparison: baseline hyperparameters were taken from their original papers or public implementations while FENRec's were tuned per dataset, so an equal tuning budget could shrink the gap.","fun_headline_variants_meta":{"raw":{"variants":["Soft future labels + enduring negatives lift SR by 6%","FENRec: future soft labels, hard negatives, +6% hits","Enduring negatives + future soft labels: +6% next-item","Future data as soft labels, durable negatives: +6% accuracy","Soft future labels + hard negatives: 6% better next-item"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001057,"raw_usage":{"total_tokens":4415,"prompt_tokens":905,"completion_tokens":3510,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":521,"completion_tokens_details":{"reasoning_tokens":3418}},"tokens_in":521,"tokens_out":3510,"duration_ms":21417,"temperature":1.0,"reasoning_tokens":3418,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T14:47:17.040983+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the main table giving all fourteen baselines the same per-dataset hyperparameter search budget as FENRec, with FENRec's own hyperparameters fixed before seeing test metrics; if the average improvement falls to the noise level, the central state-of-the-art claim is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the mixed-negatives construction (MixCSE) that inspires the enduring hard negatives."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DuoRec is the framework used in the compatibility experiments and the source of the uniformity metric."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ICSRec is the backbone system into which FENRec is integrated for the main comparison table."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CL4SRec serves both as a representative contrastive baseline and as a second framework for the +FENRec compatibility test."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ICLRec sets the intent-contrastive baseline and supplies the dataset preprocessing conventions FENRec follows."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"MVS is the label-smoothness baseline whose soft-label approach FENRec contrasts with its own future-based soft labels."}],"review_version":1}