{"id":"31880693-9b42-4504-8a13-bc2f0973b4d7","arxiv_id":"2501.15817","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Long-term Interest Clock improves streaming music recommendation by retrieving item-level past behaviors near the current time-of-day and aggregating them with time-gap-aware attention, yielding small but significant online and offline gains.","lead":"ByteDance researchers propose Long-term Interest Clock, a module for streaming music recommendation that finds a user's past interactions near the current time of day, matches them against the candidate song, and aggregates them with time-aware attention. In an online A/B test it lifted user active days by 0.122% and was deployed in the Douyin Music app.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The gains are not traced to the time-clock mechanism: LIC changes span, granularity, and time gap together, and no ablation isolates the clock component.","rationale":"I agree with the reader that the no-recency weighting in Eq. (2) is a genuine risk: same-clock behaviors from months earlier have equal retrieval weight, so long-run taste drift could inject stale interests. However, I see that as subordinate to a more basic attribution problem. The Gaussian Clock baseline differs from LIC in multiple dimensions, and no controlled ablation isolates the relative time-gap component. A standard two-stage long-term model (TWIN or SIM) uses item similarity plus absolute time; LIC's structural novelty is replacing absolute time with relative within-day time. Without a variant that turns off that component, the offline and online improvements cannot be assigned to fine-grained time perception. This is why the verdict should remain CONDITIONAL rather than ACCEPT: the requested ablation is necessary before the central claim is supported. I also note that Table 1's printed 95% confidence intervals are internally inconsistent as written (for example, Active Days lift 0.122% appears outside the bracket [-0.052%, +0.052%], and DisLike -0.683% outside [-0.586%, +0.586%]); if those brackets are literal CIs for the lifts, the statistical-significance sentence is unsupported. That should be rechecked, but it is a reporting issue separate from the attribution confound.","tokens_in":6951,"tokens_out":8723,"duration_ms":86810,"concrete_test":"On DouyinMusic-20B, run an ablation of LIC with the same Clock-GSU/Clock-ESU architecture and the same long-term item-level sequence, but set s(Delta(t_bm, t_cur)) equivalent to 0 in Eq. (2) (item-similarity-only retrieval and attention), plus a second variant that replaces the within-day gap with absolute recency, for example log(days since behavior). If either variant matches LIC's AUC/UAUC within noise, the clock/time-gap mechanism is not responsible for the reported gains; if LIC still beats both by roughly the reported +1.16% RelaImpr, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Relative to the strongest baseline (Gaussian Clock), LIC changes three variables at once: (1) long-term (about one year, tens of thousands of behaviors) versus short-term 30-day history, (2) item-level behavior embeddings versus category-level aggregated interests, and (3) the relative within-day time-gap signal s(Delta(t_bm, t_cur)) in Eq. (2), used in both Clock-GSU and Clock-ESU. The paper reports no experiment that holds (1) and (2) fixed while removing or replacing the time gap, and no long-term sequential baseline (for example, TWIN or SIM with absolute timestamps or no time feature) is included. Consequently, the +1.16% RelaImpr (AUC) and +0.122% Active Days could be produced by adding long-term, item-level behavior data rather than by the paper's claimed time-perception mechanism. That is load-bearing because the title, abstract, and contributions are specifically about the interest-clock/time-gap method. The reader's no-recency concern is a related design risk, but the missing ablation is the more direct threat: without it, the central attribution is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes Long-Term Interest Clock (LIC), a component for streaming recommendation systems that models users' within-day dynamic interests from long-term, item-level behavior sequences. LIC first retrieves, for each candidate item, the top-K behaviors whose item embedding is similar to the candidate and whose time-of-day is close to the current time (Clock-GSU, Eq. (2)), then aggregates the retrieved sequence with a time-gap-aware multi-head attention module (Clock-ESU, Eqs. (3)-(4)). The authors evaluate on the proprietary DouyinMusic-20B dataset and report offline AUC 0.6720 (+1.16% RelaImpr) and UAUC 0.6113 (+1.49%) over the base model, as well as an online A/B test with +0.122% Active Days (statistically significant) and several engagement metrics. They state that LIC has been deployed in Douyin Music.","tokens_in":7195,"tokens_out":6245,"duration_ms":61310,"significance":"The result is practically significant if the gains can be attributed to the time-perception mechanism. The online A/B test is a genuine out-of-sample check, and the reported confidence interval for Active Days (excluding zero) is a real strength, as is the stated deployment in a large industrial system. The paper also makes a useful conceptual contribution by conditioning item-level long-term behavior retrieval on time-of-day in a streaming setting. However, the current evidence does not isolate the contribution of the within-day time gap from simultaneous changes to history length and representation granularity. The absence of an ablation and of long-term sequential baselines weakens the central attribution claim, which is the main reason this manuscript needs revision.","major_comments":[{"comment":"The reported gains are not traceable to the time-gap mechanism. Relative to the strongest baseline (Gaussian Clock), LIC changes at least three factors simultaneously: long-term (about one year) versus 30-day history, item-level behavior embeddings versus category-level aggregated interests, and the relative within-day time-gap term s(Delta(t_bm,t_cur)). No experiment holds history length and representation granularity fixed while removing or replacing the time-gap signal, and no long-term sequential baseline (e.g., TWIN, SIM, or TwinV2 with absolute timestamps or without time features) is included. Consequently, the +1.16% RelaImpr in AUC and the +0.122% Active Days gain could be produced by the additional long-term, item-level behavior data rather than by the claimed time-perception mechanism. I request an ablation with s(.) set to zero or replaced by an absolute recency feature, and comparisons to long-term sequential baselines.","section":"Section 4, Table 2 and Section 3.2, Eq. (2)"},{"comment":"The DisLike row reports a relative change of -0.683% with a 95% confidence interval of [-0.586%, +0.586%], so the interval excludes zero, but the paper never states whether lower or higher DisLike is better. If lower is better, this is a statistically significant improvement; if higher is better, it is a statistically significant degradation. Since DisLike is presented as a constraint metric, the authors should state the sign convention, clarify whether the constraint is satisfied, and discuss the result explicitly.","section":"Section 4, Table 1"},{"comment":"The time-similarity term depends only on the circular within-day gap Delta(t_bm,t_cur), so a behavior from one year ago at the same clock time receives the same time weight as a behavior from yesterday. If user tastes drift seasonally or across life changes, periodic aggregation can inject outdated interests into the current-interest embedding. The paper should either justify the implicit stationarity assumption or add a recency-aware variant (e.g., multiplying by an absolute recency weight or including an absolute timestamp feature) as an ablation.","section":"Section 3.2, Eq. (2)"}],"minor_comments":[{"comment":"The offline results report single-point RelaImpr values with no confidence intervals or significance tests. Given the small differences (e.g., 0.38 percentage points of AUC between Gaussian Clock and Long-term Clock), variance estimates are needed to assess whether the offline improvements are reliable.","section":"Section 4, Table 2"},{"comment":"The method is motivated by one year of long-term behaviors, but the DouyinMusic-20B collection window is described as 8 weeks from August to September 2023. Please clarify how the one-year-long behavior sequence is constructed from logs outside this window, and report basic statistics such as the distribution of long-sequence lengths.","section":"Section 4, Datasets"},{"comment":"The first contribution bullet says 'we are the first' without the qualifier 'to our knowledge'; the related-work section already uses the softer phrasing, and the contribution bullet should be consistent.","section":"Section 1, Contributions"},{"comment":"The text says 'We apply the proposed Interest Clock on a DCN-V2-based multi-task model'; this should read 'Long-term Interest Clock' to avoid confusion with the Interest Clock baseline from Reference [10].","section":"Section 4, Online A/B Testing"},{"comment":"The sentence 'a two layer deep network' should be 'a two-layer deep network', and similar hyphenation issues should be corrected throughout the manuscript.","section":"Section 3.3"}],"recommendation":"major_revision","confidential_remarks":"This is a short industry paper from ByteDance, and the online A/B test is a genuine strength. The main risk is that the title, abstract, and contributions claim a time-perception mechanism, while the reported experiments do not isolate that mechanism from longer history and item-level granularity. I would require the ablation and the long-term sequential baselines before acceptance. The DisLike sign convention should also be resolved. The paper relies entirely on the authors' own Interest Clock baseline and a proprietary dataset; this is not a circularity problem, but an independent comparison would increase confidence in the generality of the claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a solid industrial contribution with a real deployment, but the headline attribution of the gain to the clock mechanism is not supported by the reported experiments. What's actually new: LIC replaces the coarse hourly/category-level Interest Clock with a cyclic relative time-gap (Eq. 2) and applies it in both retrieval (Clock-GSU) and attention (Clock-ESU) over item-level long-term behaviors up to a year. That combination is new relative to the cited Interest Clock, TWIN, and SIM. The online A/B is genuine out-of-sample evidence: +0.122% Active Days is statistically significant, and the system is deployed in Douyin Music. The paper is also honest about the limitations of hour embedding and gives a concrete description of the streaming instability problem.\n\nCredit where earned: the online experiment is the right kind of evidence for an industrial system paper; the retrieval/attention architecture is clearly specified, including precomputation on parameter servers, which matters for deployability; and the authors identify a real failure mode of discrete hour embeddings.\n\nSoft spots: the stress-test note is right. LIC changes three variables at once — long-term vs short-term history, item-level vs category-level behavior, and the relative time-gap feature — and there is no ablation isolating the time-gap signal. Without that, the +1.16% RelaImpr and +0.122% Active Days could be driven entirely by adding long-term item-level data, not by the time-perception mechanism that the title and abstract claim. That is load-bearing. The no-recency concern in Eq. (2) is real but secondary: time similarity depends only on time-of-day gap, so a behavior from six months ago at the same clock time is weighted as strongly as one from last week, which could inject stale interests. Worth flagging, but the missing ablation is the bigger issue. Offline results also have no error bars, only one proprietary dataset, no code, and no long-term sequential baseline with absolute timestamps. These are addressable, not fatal.\n\nSummary: this is a paper for the industrial recsys community and for researchers studying time modeling in streaming settings. It deserves a serious referee because the deployed result and the specific mechanism merit careful scrutiny, but the authors should be pushed to supply an ablation that holds history length and item granularity fixed while removing or replacing the time gap.","headline":"A real industrial deployment with a plausible time-gap mechanism, but the paper never isolates the clock component from the switch to long-term item-level behavior, so the central attribution is not yet supported.","tokens_in":7717,"tokens_out":1793,"would_cite":true,"duration_ms":18206,"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":"A streaming recommender can read current taste from behaviors that happened at the same clock time, even months ago, and adding this signal lifts ranking accuracy and user active days.","keywords":["streaming recommendation","time perception","long-term user behavior","time-gap-aware attention","interest clock","retrieval-based modeling","CTR prediction","industrial recommendation"],"falsifier":"Retrain LIC with the time-similarity term in Eq. (2) replaced by an absolute recency feature (for example, days since the behavior, or an exponential decay in $\\Delta_{\\text{days}}$) and compare AUC and UAUC on DouyinMusic-20B: if performance does not drop, the paper's central claim that cyclic clock-gap perception, not generic long-term similarity, drives the gain is refuted.","tokens_in":6746,"feed_emoji":"⏰","tokens_out":7587,"duration_ms":64173,"temperature":0.7,"pith_summary":"Today's recommendation models trained in streaming mode struggle to use time of day: at any one moment every training sample carries the same hour feature, so discrete hour embeddings cause periodic online fluctuations. This paper proposes Long-term Interest Clock (LIC), which instead computes the user's current interest embedding by retrieving, for each candidate item, the user's past item-level behaviors that happened near the current clock time and aggregating them with a time-gap-aware multi-head attention. The claim is that this retrieves fine-grained, item-level long-term interests, fixing the coarseness of the earlier Interest Clock's discrete 24-hour category-level embeddings. Offline results on a 20-billion-sample industrial music dataset show AUC 0.6720 (+1.16% relative) and UAUC 0.6113 (+1.49%), and an online A/B test reports a statistically significant +0.122% gain in user active days. If correct, LIC shows that long-term history can be reused by clock-time relevance to model within-day taste shifts in a production streaming recommender.","feed_headline":"Old listens at the same clock time improve music recommendations","feed_subtitle":"Long-term Interest Clock mines a year of behavior with time-gap attention; A/B test shows a significant active-day gain.","key_machinery":"The load-bearing object is the Long-term Interest Clock (LIC), composed of Clock-GSU and Clock-ESU. Clock-GSU performs a top-K search over up to ten thousand behaviors using a relevance score $\\alpha(\\boldsymbol{b}_m, \\boldsymbol{q}) = (W_b \\boldsymbol{b}_m \\odot W_q \\boldsymbol{q})^T / \\sqrt{d} + s(\\Delta(t_{b_m}, t_{\\mathrm{cur}}))$, where the second term is a two-layer network over the relative clock gap measured in minutes; this makes behaviors from the same time of day retrievable even if they are months old. Clock-ESU runs a time-gap-aware multi-head attention (four heads) over the retrieved subsequence, injecting $[\\Delta, \\sqrt{\\Delta}, \\Delta^2, \\log(\\Delta+1)]$ into the key/value representations so the final current-interest embedding is a fine-grained, item-level blend of what the user consumed near this clock time in the past. Precomputing the projected behavior and query embeddings and storing them in parameter servers keeps the online top-K search fast.","core_discovery":"The central claim is that fine-grained time perception in a streaming recommendation system can be achieved by making clock time the organizing key over year-long item-level behavior. Given the current request time $t_{\\mathrm{cur}}$ and a candidate item, Clock-GSU scores every long-term behavior by an item-similarity term plus a time-similarity term $s(\\Delta(t_{b_m}, t_{\\mathrm{cur}}))$ that depends only on the cyclic minute gap, and keeps the top $K=100$ behaviors. Clock-ESU then aggregates this subsequence with a four-head attention in which the relevance of each behavior again combines item similarity with the same relative-time features $[\\Delta, \\sqrt{\\Delta}, \\Delta^2, \\log(\\Delta+1)]$, producing a current-interest vector fed into the ranking model. The paper argues this is the first solution that tackles time perception through long-term behaviors in real-time streaming recommendation, and reports that it beats hour embeddings and the prior discrete Interest Clock variants offline and online.","pith_inferences":["A natural extension is to apply the same cyclic-relevance idea to other periodic contexts—day of week, season, or location-time—where relevance recurs but absolute age is secondary; the paper does not discuss these.","Because the time-similarity term ignores absolute recency, users whose tastes drift over seasons could be served stale same-time interests; a slow exponential recency decay combined with $s(\\Delta)$ is a testable modification.","The retrieved top-K behaviors could double as an explanation surface ('you listened to this around the same time before'), a transparency property the discrete hour-embedding baselines lack.","The reported offline AUC gap over Gaussian Interest Clock is 0.25 percentage points; re-running the comparison on a public lifelong-behavior dataset with a one-year span would tell whether the gain depends on scale specific to the industrial data."],"forward_implications":["Streaming recommenders can drop hour embeddings and the periodic online instability they cause; the time signal becomes a retrieval query rather than a static feature.","Users with no interaction in the exact current hour still receive a time-aware signal, because retrieval and attention smooth over nearby clock minutes within long-term history.","Long-term behavior sequences (up to a year, roughly ten thousand actions) become usable for time perception at industrial scale, with precomputed projections keeping the top-K search cheap.","Deployment in a music ranking task with significant active-day gains suggests the same two-module design transfers to other streaming ranking surfaces."],"supporting_citations":[{"why":"Prior Interest Clock method; supplies the streamed recommendation setup, DouyinMusic-20B dataset, the Naive/Adaptive/Gaussian Clock baselines, and the three limitations LIC targets.","marker":"[10]"},{"why":"TWIN two-stage interest network; provides the General Search Unit / Exact Search Unit retrieval pattern that Clock-GSU and Clock-ESU adapt.","marker":"[1]"},{"why":"Search-based lifelong behavior modeling; establishes top-K search plus target-aware attention over very long user sequences.","marker":"[4]"},{"why":"Twin V2; extends two-stage lifelong modeling to ultra-long sequences and informs the scalable top-K design for Clock-GSU.","marker":"[6]"},{"why":"DCN-V2; the base multi-task ranking model on which LIC and all offline baselines are built and on which the online A/B test runs.","marker":"[8]"}],"fun_headline_variants":["Long-term clock time fine-tunes streaming music picks","Same-minute listening history improves recommendation accuracy","Time-gap attention boosts streaming recommendation performance","Clock-based retrieval lifts long-term interest modeling","Fine-grained time perception for streaming recs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that behavior from the same clock time remains a trustworthy signal of current taste regardless of how old it is, so an 8 a.m. listen from a year ago is retrieved as eagerly as an 8 a.m. listen from last week.","fun_headline_variants_meta":{"raw":{"variants":["Long-term clock time fine-tunes streaming music picks","Same-minute listening history improves recommendation accuracy","Time-gap attention boosts streaming recommendation performance","Clock-based retrieval lifts long-term interest modeling","Fine-grained time perception for streaming recs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000494,"raw_usage":{"total_tokens":2476,"prompt_tokens":1047,"completion_tokens":1429,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":1362}},"tokens_in":663,"tokens_out":1429,"duration_ms":13247,"temperature":1.0,"reasoning_tokens":1362,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T13:55:38.119472+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain LIC with the time-similarity term in Eq. (2) replaced by an absolute recency feature (for example, days since the behavior, or an exponential decay in $\\Delta_{\\text{days}}$) and compare AUC and UAUC on DouyinMusic-20B: if performance does not drop, the paper's central claim that cyclic clock-gap perception, not generic long-term similarity, drives the gain is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"TWIN two-stage interest network; provides the General Search Unit / Exact Search Unit retrieval pattern that Clock-GSU and Clock-ESU adapt."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Search-based lifelong behavior modeling; establishes top-K search plus target-aware attention over very long user sequences."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Twin V2; extends two-stage lifelong modeling to ultra-long sequences and informs the scalable top-K design for Clock-GSU."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"DCN-V2; the base multi-task ranking model on which LIC and all offline baselines are built and on which the online A/B test runs."}],"review_version":1}