{"id":"36b14a54-df0c-43b3-b429-8d134eebb933","arxiv_id":"2608.12975","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A plug-and-play weight fusion that down-weights short dwell-time clicks improves two session-based recommendation models, and DTAMLP combines it with FFT-based filtering.","lead":"This paper finds that blending attention weights with a threshold-capped function of click-time intervals, to down-weight very short dwell times, improves session-based recommendation accuracy on TiSASRec and SR-GNN. It packages this trick with FFT-based filtering into an all-MLP model, DTAMLP, validated on Diginetica and RetailRocket, but presents both mechanisms as modest empirical findings and disclaims state-of-the-art status.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No control condition in Table 1: gains could come from reweighting itself, not from the interval signal; a randomized/inverted interval control would settle whether the sporadic-noise claim is real.","rationale":"The reader identifies the dwell-time proxy as the weakest assumption. My concern is more basic: even granting the proxy, the paper never tests whether the interval values themselves are responsible for the gain. The only contrast is fusion vs. no fusion, so the result is equally compatible with a generic reweighting/regularization effect. This is load-bearing because the paper's headline contribution is specifically that short click intervals carry exploitable sporadic-noise signal, not merely that reweighting attention can help. The proposed control experiment is cheap, uses the paper's own protocol, and would decisively separate the two explanations. I keep the verdict as CONDITIONAL/UNCHANGED because the paper's own framing is appropriately modest and the claim is falsifiable; however, if the control test showed that random or inverted intervals produce the same gains, the central explanatory claim would need to be withdrawn or substantially weakened.","tokens_in":19476,"tokens_out":5459,"duration_ms":60949,"concrete_test":"Run the exact Table 1 protocol on TiSASRec (MovieLens-1m) and SR-GNN (Diginetica) with three variants of α2: (i) original intervals as in Eq. 2; (ii) a random permutation of the same intervals within each session; (iii) inverted intervals, e.g., softmax(-min(interval, η)) or softmax(max_interval - min(interval, η)). Use the same η rule, the same β grid or learnable β, the same data splits, and at least 5 seeds. If variants (ii) or (iii) match or exceed the original gain over Origin, the interval signal is not the load-bearing component and the sporadic-noise interpretation fails; if only original intervals improve, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The plug-and-play result is presented as the core evidence for Insight 1 (sporadic noise), but the only comparison in Table 1 and Section 3.2.1 is \"weight fusion vs. unmodified backbone.\" The fusion operation in Eq. 3 is a convex combination of the model attention α1 and α2 = softmax(min(interval, η)). This is a monotone reweighting that always down-weights the shortest intervals, regardless of whether those intervals actually carry sporadic noise. Nothing in the design or experiments separates the contribution of the interval values from the contribution of any smooth reweighting or regularization of the attention distribution. In particular, there is no control with shuffled intervals, with inverted intervals (softmax over -interval), or with a position-based weight of the same functional form. Without such a control, the observed gains are compatible with the alternative hypothesis that any soft reweighting of attention helps, e.g., by reducing overconfident attention weights. If that alternative is true, the central claim that the threshold-capped time-interval weight is the active ingredient is not established, even though the engineering trick might still produce gains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper makes two empirical claims and combines them in DTAMLP. The first, presented as the central result, is that click-time intervals in session-based recommendation contain 'sporadic noise': very short intervals often reflect accidental or low-preference clicks, so fusing a model's attention weight with a threshold-capped softmax over intervals (Eq. 3; Eqs. 15–16) improves accuracy when plugged into TiSASRec and SR-GNN (Table 1). The second is a conjecture, explicitly labeled as unproven, that FFT-based filtering helps because it separates entangled 'preference noise'; DTAMLP integrates weight fusion with an FFT-Transformer and a representation-consistent embedding design, and is evaluated on Diginetica and RetailRocket (Tables 3 and 4), with ablations and parameter sensitivity.","tokens_in":19722,"tokens_out":7833,"duration_ms":75771,"significance":"If the plug-and-play weight fusion result were established, it would be practically useful: it is a near-cost-free modification applicable to widely used SBR backbones, and the paper's framing makes the claim falsifiable. The paper is also commendably honest about the status of its second claim: the preference-noise explanation is explicitly an interpretation, not a mechanism, and the system-level results are not presented as state of the art. The main weakness is that the central empirical claim is not yet supported at the standard needed for publication: Table 1 lacks statistical replication and, more importantly, lacks control conditions that separate the interval signal from generic attention reweighting. The DTAMLP results depend on tuning choices that are not shown to be stable across datasets. Because the required evidence is obtainable within the scope of the manuscript, I regard this as a major-revision situation rather than a rejection.","major_comments":[{"comment":"The claimed evidence for sporadic noise is incomplete. The fusion operation α = (1−β)α1 + β softmax(min(interval,η)) is a monotone reweighting that always down-weights the shortest intervals, regardless of whether those intervals actually carry sporadic noise. The experiments compare only “weight fusion vs. unmodified backbone,” so the gains in Table 1 are also compatible with the alternative hypothesis that any soft reweighting of attention helps (e.g., by reducing overconfident attention weights). Please add control conditions with shuffled intervals, inverted intervals (softmax over −interval), and a position-based weight of the same functional form, and report whether the gains persist. Without these controls, the central claim that the threshold-capped time-interval weight is the active ingredient is not established.","section":"§3.2.1, Eq. (3), Table 1"},{"comment":"All accuracy numbers are single-run metrics with no standard deviations, confidence intervals, or significance tests. The TiSASRec gains in Table 1 (NDCG@10 from 56.8163 to 57.1112 at β=0.1, and HR@10 from 79.3377 to 80.1324) are small and could easily lie within run-to-run noise; the same concern applies to many differences in Tables 3 and 4. Please report means and standard deviations over at least 3–5 seeds and, for the plug-in comparisons, a paired significance test over sessions. This is necessary to support the abstract's phrase “consistent accuracy gain.”","section":"Table 1; Tables 3–4"},{"comment":"Hyperparameters intervalmax, margin, dropout, and embedding size are tuned on RetailRocket and then RetailRocket results are reported in Table 4. If this tuning used the test set, the reported numbers are optimistic and the procedure is circular with respect to robustness claims. Please state explicitly which split was used for tuning, fix hyperparameters before evaluating the test set, and show that the chosen values (or the qualitative conclusion) are stable across datasets or over a range of settings.","section":"§5.3.3; Tables 3–4"},{"comment":"The premise that the interval between two adjacent clicks is a usable proxy for dwell time, and that very short intervals predominantly indicate low-preference sporadic clicks, is asserted via an anecdotal shopping example. No distribution of intervals, no behavioral validation, and no robustness check with alternative interval definitions are provided. If short intervals often reflect rapid but genuine exploration, down-weighting them would reduce real preference signal. Please either test this premise directly (e.g., relate intervals to next-click relevance) or show that the plug-in gain is insensitive to the interval definition.","section":"§3.2.1"}],"minor_comments":[{"comment":"The heading contains a typo: “Promblem Formulation” should be “Problem Formulation.”","section":"§3.1"},{"comment":"The variant label “w/o None” is confusing; please use “Full model” or “No component removed.”","section":"§5.3.2"},{"comment":"Figures 6 and 7 are not rendered correctly in the submitted version: the panels contain raw UTF strings (e.g., “/uni00000013/...”), and the panel labels do not match the text, which describes tuning of intervalmax, margin, dropout, and embedding size. Please regenerate the figures.","section":"Figures 6–7"},{"comment":"The complexity comparison states O(d log d) for the FFT-Transformer but does not account for the DNN, time-embedding, and prediction components; the reported training time is not tied to a reproducible configuration. Please clarify the scope of the complexity claim.","section":"§5.4"},{"comment":"The phrase “consistent accuracy gain” overstates what single-run results can support; please qualify it as gains observed in the reported runs.","section":"Abstract"},{"comment":"No code or data release is mentioned; since the central contribution is a plug-and-play modification, providing code (or pseudocode for the weight-fusion module) would greatly aid verification.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest about its limitations, and the central claim is falsifiable in principle, but the missing control conditions and lack of statistical replication are load-bearing. I would be willing to accept after the authors add the proposed controls and either fix the tuning protocol or demonstrate robustness. The unusual presentation (figures, the footnote about the paper's history) suggests a nonstandard submission, but I do not see that as a scientific issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: the weight-fusion module is a genuinely cheap idea – blend a model's attention weight with a threshold-capped softmax over click intervals – and the demonstration on two architecturally different backbones (TiSASRec, SR-GNN) is easy to believe as a useful trick. But the paper never actually shows that the interval values are the active ingredient. Eq. 3 is a convex combination that always down-weights the shortest intervals; any smooth reweighting of attention might give the same gain. There's no shuffled-interval or inverted-interval control, so the 'sporadic noise' interpretation is not established. That's not fatal for the engineering claim, but it is the difference between a trick and a finding.\n\nWhat the paper does well: it is clearly scoped and unusually honest. It labels the frequency-domain 'preference noise' story as an interpretive conjecture, disclaims SOTA status, and points to the plug-in verification as its most direct contribution. The interval-capped softmax fusion is not in the cited TiSASRec or SR-GNN work, as far as I can tell, so the novelty claim is reasonable. The ablation structure is sensible, though the numbers are weak support.\n\nThe soft spots, in order: (1) no control condition, as above; (2) all metrics are single-run with no standard deviations or significance tests – the TiSASRec NDCG gains in Table 1 are small enough to be noise; (3) hyperparameters including intervalmax and margin are tuned on RetailRocket and then the tuned numbers are reported as the main result, which is a form of test-set leakage; (4) the interval-as-dwell-time proxy is asserted from a shopping anecdote, with no distribution analysis or robustness check; (5) no code released, so independent verification means reimplementing.\n\nThe preference-noise conjecture is clearly labeled as speculation, so I don't fault them for it. I'd rather see the paper spend its space on a proper control for the weight fusion.\n\nWho should read this: anyone working on time-aware or GNN-based session recommendation who wants a quick trick to try. It deserves a serious referee, because the central claim is cheap to verify and the write-up is honest enough that a revision with controls and multi-seed runs could make it solid.","headline":"A cheap, plausible plug-in trick for time-aware recommenders, reported honestly but with evidence that doesn't yet isolate the interval signal.","tokens_in":20267,"tokens_out":2288,"would_cite":false,"duration_ms":24326,"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 central claim is that down-weighting clicks with very short time intervals, via a plug-and-play weight fusion module, consistently improves time-aware and GNN-based session recommenders.","keywords":["session-based recommendation","sporadic noise","time interval","weight fusion","frequency-domain filtering","FFT","preference noise","MLP recommender"],"falsifier":"Take the same plug-in weight fusion and either reverse its direction, up-weighting the shortest intervals, or apply it to a dataset with true dwell-time labels: if reversed weighting also helps, or if short-interval clicks are frequently followed by a same-item re-click, purchase, or long subsequent view, then the central claim that short gaps are sporadic noise fails.","tokens_in":19233,"feed_emoji":"🛒","tokens_out":8547,"duration_ms":75572,"temperature":0.7,"pith_summary":"This paper tries to establish a small, practical claim about session-based recommendation: models that use click timestamps treat every gap between clicks as equally meaningful, but very short gaps often mark accidental, low-preference clicks. It proposes a plug-and-play weight fusion module that blends a model's existing attention weight with a threshold-capped softmax weight over click intervals, so clicks with tiny intervals are down-weighted. Applied without touching the backbone, the module improves TiSASRec on MovieLens-1m and SR-GNN on Diginetica, which the paper calls its most directly verifiable result. The paper also offers a speculative explanation for why frequency-domain filtering helps, and packages both ideas in an all-MLP model, DTAMLP, validated on Diginetica and RetailRocket. The system-level accuracy numbers are presented as supporting evidence, not as a state-of-the-art claim.","feed_headline":"Down-weighting super-fast clicks lifts session recommenders","feed_subtitle":"A plug-in blend of attention and capped click-interval weights improves TiSASRec and SR-GNN without retraining.","key_machinery":"The load-bearing object is the weight fusion identity: the final item weight is a blend of the model's own attention weight and a softmax over capped time intervals, $\\alpha_2 = \\mathrm{softmax}(\\min(\\mathrm{interval}, \\eta))$, combined as $\\alpha = (1-\\beta)\\alpha_1 + \\beta\\alpha_2$. The cap $\\eta$ is the upper limit beyond which longer gaps no longer add signal, so only suspiciously short intervals are punished. In DTAMLP the same recipe appears as Eqs. 15-16 with a learnable $\\beta$, alongside an FFT-Transformer, $\\tilde{F}^l = \\mathcal{F}^{-1}(W \\odot \\mathcal{F}(F^l))$, that filters item embeddings in the frequency domain, plus a representation-consistent session embedding adapted from prior work. The weight fusion carries the paper's central claim; the FFT module carries its interpretive conjecture.","core_discovery":"On its own terms, the central discovery is that click-time intervals carry exploitable signal about sporadic noise: a very short gap between two adjacent clicks tends to follow a click made out of curiosity or misleading appeal, not genuine preference. The paper formalizes this with a capped interval weight $\\alpha_2 = \\mathrm{softmax}(\\min(\\mathrm{interval}, \\eta))$ and fuses it with the model's attention weight $\\alpha_1$ via $\\alpha = (1-\\beta)\\alpha_1 + \\beta\\alpha_2$; the blend coefficient $\\beta$ can be fixed or learnable. This one equation, inserted into TiSASRec and SR-GNN with no other architectural change, produces consistent gains, including SR-GNN's MRR@20 rising from 15.83 to 18.01 on Diginetica. The paper is explicit that the second claim, that Fourier-domain filtering helps because it separates entangled 'preference noise,' is an interpretive conjecture, not a proven mechanism.","pith_inferences":["Beyond the paper: a direct extension would be to test the same fusion across more backbones and datasets and measure whether the optimal threshold $\\eta$ tracks session length or domain, which the paper does not report.","Beyond the paper: if short intervals genuinely mark sporadic noise, then the module's down-weighting should agree with true dwell-time labels from logs such as scroll depth or video watch time; the paper only uses click gaps.","Beyond the paper: the preference-noise conjecture could be probed by inspecting learned FFT filters and checking whether items sharing a filter response cluster by interpretable attributes, a test the paper leaves open.","Beyond the paper: because the module needs no architectural change, it could operate as a post-hoc scoring adjustment on a deployed model, an operational consequence the paper does not develop."],"forward_implications":["Time-aware recommenders can gain accuracy by down-weighting clicks separated by very short intervals, without retraining the backbone.","The fusion can be added to attention-based and GNN-based models even when they have no explicit time-awareness, as the SR-GNN result shows.","Ablations on DTAMLP indicate that time-interval weight fusion and frequency-domain filtering improve accuracy independently and complementarily.","The gain is larger on ranking-sensitive metrics than on recall, so the intervention mostly reorders candidates rather than retrieving a different set of relevant items."],"supporting_citations":[{"why":"Supplies TiSASRec, the time-aware self-attention backbone used in the plug-and-play weight fusion test.","marker":"Li et al., 2020"},{"why":"Supplies SR-GNN, the GNN-based backbone with no explicit time-awareness that also improves when weight fusion is added.","marker":"Wu et al., 2019"},{"why":"Reports the under-explained frequency-domain filtering improvement in FMLP-Rec that motivates the preference-noise conjecture.","marker":"Zhou et al., 2022"},{"why":"Provides the representation-consistent session embedding and robust distance measure that DTAMLP adopts.","marker":"Hou et al., 2022"},{"why":"Defines SASRec, a baseline in the main comparison, and the Transformer-style skip-connection and normalization conventions the FFT-Transformer block follows.","marker":"Kang & McAuley, 2018"}],"fun_headline_variants":["Short clicks are noise: fuse capped time weights for better sessions","Capping click intervals boosts session recommenders","Down-weighting accidental clicks improves session models","A weight blend that drowns out rapid-click noise"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on treating the interval between adjacent clicks as a proxy for how long the user actually looked at the item, so that a very short interval marks a low-preference sporadic click.","fun_headline_variants_meta":{"raw":{"variants":["Short clicks are noise: fuse capped time weights for better sessions","Capping click intervals boosts session recommenders","Down-weighting accidental clicks improves session models","A weight blend that drowns out rapid-click noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1377,"prompt_tokens":1005,"completion_tokens":372,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":311}},"tokens_in":621,"tokens_out":372,"duration_ms":4056,"temperature":1.0,"reasoning_tokens":311,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:22:11.250738+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the same plug-in weight fusion and either reverse its direction, up-weighting the shortest intervals, or apply it to a dataset with true dwell-time labels: if reversed weighting also helps, or if short-interval clicks are frequently followed by a same-item re-click, purchase, or long subsequent view, then the central claim that short gaps are sporadic noise fails.","supporting_citations":[],"review_version":1}