{"id":"b37304da-8110-480b-b634-453a1b4cf246","arxiv_id":"2505.08471","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A recommendation model that encodes user interest life-cycle states from activity histograms and uses them to gate expert networks, reporting small online gains on the Lofter app.","lead":"This paper introduces a recommendation model that tracks how each user interest grows, stabilizes, and fades, and uses that life-cycle signal to rerank content. The model is already live on the Lofter app, with small measured gains in click, conversion, and reading time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Online A/B gains are reported as 'significant' without significance tests, confidence intervals, or control-group details; the 0.38% CTR and 1.04% CVR lifts may be within normal fluctuation, so the deployment claim is not yet established.","rationale":"The reader's weakest_assumption focuses on feedback circularity: exposure histograms are outputs of the recommender's own policy, so DILN might create the emergent-interest signals it then exploits. That is a real interpretive concern, but it is not the most load-bearing one for the central claim, because the offline experiment on KuaiRand uses randomly exposed videos, which breaks the exposure-policy feedback loop and still shows GAUC gains for DILN. This suggests the method captures something beyond a pure policy artifact, although the offline result is itself reported without variance or significance. The reader's rationale already mentions 'no significance tests' as one of several thin spots, so my concern partially overlaps with the reader's assessment but selects a different element as the primary load-bearing risk. The most direct threat to the paper's headline 'deployed and effective' claim is that the online A/B lifts of 0.38% CTR and 1.04% CVR are small relative to typical platform noise and are presented with no statistical evidence. Because the reader's verdict is already CONDITIONAL and explicitly asks for statistical detail, my concern does not move the verdict; it sharpens the condition under which acceptance would be justified.","tokens_in":7118,"tokens_out":9169,"duration_ms":101749,"concrete_test":"Ask the authors for the A/B experiment's per-day metric time series and the 95% confidence intervals (or p-values) for CTR, CVR, and duration, computed on the user-level data with the stated randomization unit and a standard multiple-comparison correction. If all three intervals exclude zero, the deployment claim stands; if any interval contains zero, the central claim is unsupported and the paper should be treated as an unverified industrial case study rather than a demonstrated deployment success.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DILN is a working industrial upgrade, supported by the online A/B results in Section 4.3: +0.38% CTR, +1.04% CVR, and +0.25% duration over the production model after three weeks on more than 20% of users. Yet the paper gives no p-values, confidence intervals, standard deviations, daily time series, or any description of control-group allocation, randomization unit, or overlap handling. In large recommender platforms, daily metric fluctuations are often on the order of tenths of a percent, and three-week aggregates can be affected by novelty effects or concurrent launches. Without distributional evidence, the word 'significant' is an assertion rather than a demonstrated result. This is the most load-bearing weakness because if the lifts are indistinguishable from noise, the deployment statement and the exposure-shift analysis in Table 2 lose their empirical ground, regardless of the architectural plausibility of DILN. The offline GAUC gains on KuaiRand provide supporting evidence, but they too are reported without variance or significance tests and use a different metric and a one-day test split. No code, configuration, or feature definitions for the predefined interest categories are provided, so the offline results cannot be independently reproduced.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Deep Interest Life-cycle Network (DILN) for industrial recommendation ranking. DILN consists of an Interest Life-cycle Encoder Module (ILEM), which builds per-candidate activity histograms over recent exposure/click/interaction events weighted by search relevance, encodes them with a CNN, and discretizes them with vector quantization, and an Interest Life-cycle Fusion Module (ILFM), which injects the resulting life-cycle embedding into an MMOE-style ranking model through a feature recalibrator and neural fusion units. The authors report offline GAUC improvements on the public KuaiRand dataset and on an industrial Lofter dataset, and an online A/B test on Lofter over three weeks on more than 20% of users yielding +0.38% CTR, +1.04% CVR, and +0.25% duration improvements, together with a dispatch analysis showing increased exposure of emergent interests. The paper states that DILN has been deployed on the Lofter App.","tokens_in":7354,"tokens_out":4568,"duration_ms":45106,"significance":"The contribution is practical and industrially relevant: DILN is a lightweight module that can be attached to an existing ranking model, and the ablation SIM vs SIM+ILEM vs DILN in Table 1 is directionally consistent with the module's usefulness. The use of a public dataset (KuaiRand) and the deployment statement are notable strengths, as is the dispatch analysis in Table 2, which gives insight into the mechanism of the method. However, the quantitative claims currently rest on point estimates without statistical inference, and the features built from exposure logs raise a potential feedback-loop concern that could affect the life-cycle interpretation. If the online gains survive proper significance testing and the life-cycle clusters are validated against randomized exposure data, this would be a solid industrial contribution; as presented, the evidence is suggestive rather than conclusive.","major_comments":[{"comment":"The online A/B results are the central evidence for the deployment claim, but the paper reports only point estimates (+0.38% CTR, +1.04% CVR, +0.25% duration) with no p-values, confidence intervals, standard errors, daily or weekly time series, or description of randomization unit, control-group size, overlap handling, and metric definitions. Because metric fluctuations of tenths of a percent are common in large recommender systems, the word 'significant' is an unsupported assertion. The authors should describe the A/B setup and report bootstrap confidence intervals or a significance test, and ideally a daily time series to rule out novelty or concurrent-launch effects.","section":"Section 4.3, Online Result"},{"comment":"The histograms that define interest life-cycle are constructed from exposure logs, but exposure is itself an output of the previous ranking policy. A model that learns to expose more emergent-interest content will increase future exposure histogram mass for emergent interests, making the VQ cluster assignments and the dispatch shift in Table 2 partly self-fulfilling. To support the claim that the clusters encode user interest phases rather than policy artifacts, the authors should compare cluster and histogram distributions on randomly exposed data (for example, KuaiRand's random-exposure setup, or a randomized traffic holdout) against logged exposure data; a substantial difference would indicate that the features capture the recommender's own prior behavior.","section":"Section 3.1.1, Feature engineering; Table 2"},{"comment":"The offline comparison uses only a single baseline architecture (MMOE+SIM) and reports GAUC on a single test day for both datasets, with no confidence intervals, multiple seeds, or significance tests. The claimed 'significant improvement' in Section 4.2 is therefore not statistically supported, and the one-day test window is too short to assess GAUC stability. In addition, the pre-defined interest categories used in Figure 4 are never defined, so the alignment between learned VQ clusters and business categories cannot be checked. Please report multiple test days or seed-level variance, define the categories, and ideally add at least one stronger recent lifelong-interest baseline to situate the gains.","section":"Section 4.1, Experimental Settings; Table 1"}],"minor_comments":[{"comment":"The scaling factor gamma is fixed to 2 with no ablation or sensitivity analysis; please clarify whether this value was tuned and how sensitive the results are to it.","section":"Section 3.2.1, Eq. (2)"},{"comment":"'a emergent phase' should be 'an emergent phase'.","section":"Abstract"},{"comment":"'Exposure Active Histogram' should be 'Exposure Activity Histogram'.","section":"Figure 2 caption"},{"comment":"The baseline is described only as 'our latest production model'; name the model and version so that readers can assess the comparison.","section":"Section 4.3"},{"comment":"The CNN kernel sizes and filter counts are stated, but no rationale or ablation is given for this specific architecture.","section":"Section 3.1.2"},{"comment":"The stop-gradient notation sg is used without definition; please define it in the text.","section":"Section 3.1.3, Eq. (1)"},{"comment":"The units and denominators of the CTR/CVR changes within each interest category are unclear; state whether these are relative or absolute percentage-point changes and what the denominator is for each row.","section":"Table 2"},{"comment":"The claim of being 'the first to introduce interest life-cycle features' is broader than what can be verified from the cited related work; please narrow the novelty claim or add a comparison with interest-evolution methods.","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for an industrial-track recommender-systems venue and the proposed module is plausible. My main reservations are the absence of any statistical inference for the online A/B results and the potential feedback loop between exposure-based features and the model's own policy. If the authors can provide significance testing, confidence intervals, and a randomized-exposure validation of the life-cycle clusters, I would be willing to support acceptance; without those, the central deployment claim is not fully established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a competent industrial case study with a genuinely new combination: activity histograms over GSU-retrieved behaviors, CNN encoding, VQ clustering into life-cycle codes, and MMOE-style gating. The ablation SIM vs SIM+ILEM vs DILN is monotonic, and the offline GAUC gains on KuaiRand give some independent support. The exposure-shift analysis in Table 2 is a nice attempt to show the mechanism, not just the metric lift.\n\nThe soft spots are mostly evidentiary, not architectural. The online A/B section reports +0.38% CTR, +1.04% CVR, and +0.25% duration as \"significant\" but gives no p-values, confidence intervals, daily time series, randomization unit, or overlap handling. The baseline is just \"our latest production model\". On a large platform, those lifts could easily sit inside normal daily fluctuation, so the deployment claim is not yet established. The offline gains are also reported without variance, and the gap between ~4.5% relative GAUC lift and 0.38% CTR lift is unexplained. The exposure histograms are built from the recommender's own logs, so the VQ clusters could partly encode policy artifacts rather than true interest phases; Figure 4's mapping to business categories is interpretive, not validated. The contribution claim of being \"the first to introduce interest life-cycle features\" is too strong without a broader literature scan. No code or configuration is provided, so the specific numbers are not independently checkable.\n\nThat said, these are the usual weaknesses of a short industrial paper, not fatal flaws. The method is clearly described, the approach is plausible, and the offline ablation plus the directional consistency of the online results suggest the module does something. If a referee pushes for statistical detail and a clearer baseline description, the paper would be much stronger.\n\nI would send this to peer review. It deserves referee time as an industrial contribution to user interest modeling, with the expectation of heavy revision on the experimental reporting. I would not cite it in my own work until the statistical evidence is shored up, but I would put it on a reading group list for the architectural idea.","headline":"Plausible industrial module for modeling interest phase; the architecture is new enough to be worth a referee, but missing significance tests and baseline details leave the headline deployment claim under-supported.","tokens_in":720,"tokens_out":920,"would_cite":false,"duration_ms":25635,"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":"This paper claims that a recommender can read the life-cycle phase of each user interest — emergent, stable, or declining — from activity histograms and use those phases to rerank items, yielding higher CTR, CVR, and duration in production.","keywords":["recommendation system","user interest modeling","interest life-cycle","Deep Interest Life-cycle Network","multi-task learning","vector quantization","CTR prediction","online A/B testing"],"falsifier":"Design an A/B experiment in which one serving arm deliberately over-exposes declining-interest items for a week, then retrain DILN on that arm's logs and check whether the learned VQ clusters reassign those items to the 'declining' cluster more often than in a control arm. If cluster assignments move with the policy, the life-cycle features are policy artifacts; if they stay put, the features track the user's interest.","tokens_in":6879,"feed_emoji":"📈","tokens_out":8476,"duration_ms":77401,"temperature":0.7,"pith_summary":"User interests move through emergent, stable, and declining phases, and the paper argues that recommendation models that ignore this life-cycle waste exposure on fading interests and fail to capture rising ones. To close that gap, the authors propose DILN, which builds per-candidate activity histograms from exposure, click, and interaction logs, encodes them into a dense life-cycle vector, and quantizes that vector into clusters that represent interest phases. Those life-cycle features are injected into a multi-task ranking model through a feature recalibrator and gated expert layers, so that different phases activate different experts for different objectives. The authors report offline gains on KuaiRand and an industrial dataset plus an online A/B test on Lofter with +0.38% CTR, +1.04% CVR, and +0.25% duration per user, with exposure shifting toward emergent interests. If these results hold, modeling the life-cycle is a simple, deployable upgrade to existing ranking models.","feed_headline":"Interest life-cycle model lifts CTR, CVR, and duration on Lofter","feed_subtitle":"Activity histograms classify interests as emergent, stable, or declining, shifting exposure toward rising interests.","key_machinery":"The central mechanism is the interest life-cycle vector built from activity histograms. For each candidate item and each action type (exposure, click, interaction), a General Search Unit retrieves the N most relevant past behaviors; summing relevance scores per active day produces a length-K histogram of activity intensity. A 1-D CNN encodes the histogram, and a vector-quantization codebook of ten clusters turns the encoded vector into a discrete life-cycle label. That label drives two lightweight gates in the ranking model: a feature recalibrator that re-weights the shared-bottom features and neural fusion units that rescale each expert's hidden layers. The VQ label is what lets 'emergent' versus 'declining' contexts activate different experts for different tasks, which is the mechanism carrying the reported gains.","core_discovery":"DILN's central claim is that the trajectory of a user's activity toward a candidate item — ramping up, holding steady, or winding down — is a learnable signal that existing sequential-interest models leave unused. The Interest Life-cycle Encoder Module constructs, for each candidate item, histograms of recent exposure, click, and interaction activity, compresses them with a 1-D convolutional network, and maps them through a vector-quantization codebook of ten clusters, each acting as a life-cycle phase label. The Interest Life-cycle Fusion Module then uses that label to rescale input features and expert hidden states inside an MMOE multi-task model, so that emergent, stable, and declining phases get different expert behavior for CTR, CVR, and duration objectives. Against a SIM-based production baseline, the authors find consistent GAUC gains offline and, in a three-week A/B test on Lofter, significant online gains plus a dispatch shift of +5.11% impression share toward emergent interests and -4.4% away from declining interests.","pith_inferences":["The authors do not test whether the learned life-cycle phases are stable under a change of serving policy: because histograms are built from the recommender's own exposure and click logs, a policy that over-serves declining interests could manufacture the 'declining' label. One test would rerun DILN with forced random exploration and compare cluster assignments.","A natural extension is to use the VQ cluster as a structured control variable for exposure fairness or content auditing, since it gives every candidate item a discrete, interpretable life-cycle label.","The larger CVR gain (+1.04%) than CTR gain (+0.38%) is a hint the life-cycle gating helps most at the later decision stage; a follow-up could test whether the effect concentrates in users whose interest recently emerged."],"forward_implications":["Adding ILEM to the SIM baseline improves offline GAUC on both KuaiRand and the industrial dataset, and adding ILFM improves it further, so interest life-cycle features act as a plug-in upgrade to an existing ranking model.","In production, DILN shifts impression share by +5.11% toward emergent interests, -4.4% away from declining interests, and -2.17% away from unexplored interests, while long-term interests hold steady.","Emergent interests show the highest task efficiency despite sparse historical behaviors, so detecting the emergent phase early is where the largest untapped ranking gains lie.","The VQ cluster assignments are interpretable: Unexplored interests primarily activate cluster 0 while emergent interests primarily activate clusters 3 and 4, which means the cluster label can serve as a continuous monitor of interest-phase distribution.","A three-week online A/B test on over 20% of users shows +0.38% CTR, +1.04% CVR, and +0.25% duration per user against the production model."],"supporting_citations":[{"why":"Supplies the SIM baseline with lifelong search-based interest retrieval; DILN is built on SIM and compared against it in Table 1.","marker":"[13]"},{"why":"Provides the MMOE multi-task expert architecture that the Interest Life-cycle Fusion Module modifies with recalibration and gating.","marker":"[12]"},{"why":"Supplies the vector-quantization technique behind the Life-Cycle VQ Cluster that discretizes histograms into life-cycle labels.","marker":"[14]"},{"why":"Provides the KuaiRand public dataset used for offline evaluation and comparison.","marker":"[8]"},{"why":"Establishes convolutional sequence modeling as the basis for the histogram encoder in ILEM.","marker":"[1]"}],"fun_headline_variants":["Interest life-cycle signal lifts CTR, CVR, and duration on Lofter","Recsys learns interest rise and fall, shifts exposure to winners","Quantizing interest phases boosts recsys: +1% CVR, +0.38% CTR","DILN: model interest life cycles to boost engagement metrics","Interest emergence vs decline: a learnable recsys signal"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the exposure, click, and interaction histograms measure the user's underlying interest phase rather than the recommender's own prior serving policy, since exposure logs are themselves outputs of the ranking system.","fun_headline_variants_meta":{"raw":{"variants":["Interest life-cycle signal lifts CTR, CVR, and duration on Lofter","Recsys learns interest rise and fall, shifts exposure to winners","Quantizing interest phases boosts recsys: +1% CVR, +0.38% CTR","DILN: model interest life cycles to boost engagement metrics","Interest emergence vs decline: a learnable recsys signal"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000272,"raw_usage":{"total_tokens":1656,"prompt_tokens":992,"completion_tokens":664,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":608,"completion_tokens_details":{"reasoning_tokens":568}},"tokens_in":608,"tokens_out":664,"duration_ms":6315,"temperature":1.0,"reasoning_tokens":568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:54:10.094116+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Design an A/B experiment in which one serving arm deliberately over-exposes declining-interest items for a week, then retrain DILN on that arm's logs and check whether the learned VQ clusters reassign those items to the 'declining' cluster more often than in a control arm. If cluster assignments move with the policy, the life-cycle features are policy artifacts; if they stay put, the features track the user's interest.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the KuaiRand public dataset used for offline evaluation and comparison."}],"review_version":1}