Long-Term Interest Clock: Fine-Grained Time Perception in Streaming Recommendation System

Feng Zhang; Guanyu Jiang; Jingwu Chen; Xiao Yang; Yongchun Zhu; Zuotao Liu

REVIEW 3 major objections 5 minor 1 cited by

Long-Term Interest Clock: Fine-Grained Time Perception in Streaming Recommendation System

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2501.15817 v1 pith:W3CT4G62 submitted 2025-01-27 cs.IR cs.AI

Yongchun Zhu , Guanyu Jiang , Jingwu Chen , Feng Zhang , Xiao Yang , Zuotao Liu This is my paper · ORCID

classification cs.IRcs.AI

keywords streamingrecommendationtimeperceptionlong-termuserbehaviortime-gap-awareattentioninterestclockretrieval-basedmodelingCTRpredictionindustrial

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

Credit 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.

Soft 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.

Summary: 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.

Referee Report

3 major / 5 minor

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.

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 (3)

[Section 4, Table 2 and Section 3.2, Eq. (2)] 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 4, Table 1] 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 3.2, Eq. (2)] 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.

minor comments (5)

[Section 4, Table 2] 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 4, Datasets] 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 1, Contributions] 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 4, Online A/B Testing] 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 3.3] The sentence 'a two layer deep network' should be 'a two-layer deep network', and similar hyphenation issues should be corrected throughout the manuscript.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: LIC's reported gains are externally benchmarked; the only self-citations supply baselines and data, not the result.

full rationale

The paper's derivation chain is a modeling construction: define current-interest embedding as a retrieval over long-term behaviors with time-gap-aware attention (Eqs. 2-4), train with cross-entropy, and evaluate by online A/B and offline AUC/UAUC. The central claims (+0.122% Active Days, +1.16% RelaImpr) are measured against baselines rather than derived from them. The self-citation to Zhu et al. [10] provides the Interest Clock baselines and the DouyinMusic-20B dataset; these are comparison and development resources, not load-bearing premises that force the improvement. The absence of an ablation isolating the clock component is an attribution/correctness concern, not circularity: no reported metric is defined in terms of the clock term, and no fitted parameter is renamed as a prediction. No equation in the paper reduces to its own input by construction.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method introduces no new physical entities. Its central claim rests on two domain assumptions (stable daily rhythms, hour-embedding instability) and on hand-chosen time features and hyperparameters that are not ablated. The main load-bearing premise is the stability of intra-day preferences over long histories, since retrieval ignores absolute recency.

free parameters (4)

Top-K retrieval length K = 100
Chosen without reported sensitivity analysis (Section 3.2). Affects how many long-term behaviors reach attention.
Number of attention heads = 4
Set in Section 3.3 without ablation; controls capacity of time-gap-aware attention.
Time-gap feature expansion = [Δ, √Δ, Δ², log(Δ+1)] with hidden size 8
Hand-designed input to s(·) in Eq. (2); no justification or comparison to alternatives.
Long-term behavior sequence length = about 10,000
Section 3.2 states search over 'ten thousand length of user behaviors'; choice not motivated.

assumptions (3)

domain assumption User interests have stable intra-day periodic patterns that persist across the long behavior history (about one year).
Core premise for retrieving behaviors by clock-time only with no recency decay in Eq. (2). If daily rhythms drift, stale same-time behaviors mislead the current-interest embedding.
domain assumption Discrete hour embeddings cause instability and periodic online patterns in streaming recommendation training.
Adopted from [10] and stated in Section 1. It motivates replacing hour buckets; not independently tested in this paper.
ad hoc to paper The model can learn useful intra-day periodicity from features [Δ, √Δ, Δ², log(Δ+1)] through a two-layer network s(·).
No theoretical or empirical justification for this specific feature map; if the network cannot represent the needed cyclic structure, the time-similarity term fails.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Long-Term Interest Clock: Fine-Grained Time Perception in Streaming Recommendation System." pith.science (2026). https://pith.science/paper/W3CT4G62

@misc{pith2026250115817,
  author       = {Pith},
  title        = {Pith review of: Long-Term Interest Clock: Fine-Grained Time Perception in Streaming Recommendation System},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3CT4G62}},
  note         = {Machine review of arXiv:2501.15817}
}

read the original abstract

User interests manifest a dynamic pattern within the course of a day, e.g., a user usually favors soft music at 8 a.m. but may turn to ambient music at 10 p.m. To model dynamic interests in a day, hour embedding is widely used in traditional daily-trained industrial recommendation systems. However, its discreteness can cause periodical online patterns and instability in recent streaming recommendation systems. Recently, Interest Clock has achieved remarkable performance in streaming recommendation systems. Nevertheless, it models users' dynamic interests in a coarse-grained manner, merely encoding users' discrete interests of 24 hours from short-term behaviors. In this paper, we propose a fine-grained method for perceiving time information for streaming recommendation systems, named Long-term Interest Clock (LIC). The key idea of LIC is adaptively calculating current user interests by taking into consideration the relevance of long-term behaviors around current time (e.g., 8 a.m.) given a candidate item. LIC consists of two modules: (1) Clock-GSU retrieves a sub-sequence by searching through long-term behaviors, using query information from a candidate item and current time, (2) Clock-ESU employs a time-gap-aware attention mechanism to aggregate sub-sequence with the candidate item. With Clock-GSU and Clock-ESU, LIC is capable of capturing users' dynamic fine-grained interests from long-term behaviors. We conduct online A/B tests, obtaining +0.122% improvements on user active days. Besides, the extended offline experiments show improvements as well. Long-term Interest Clock has been integrated into Douyin Music App's recommendation system.

Figures

Figures reproduced from arXiv: 2501.15817 by the authors.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

RecoWorld: Building Simulated Environments for Agentic Recommender Systems
cs.IR 2025-09 conditional novelty 5.0 of 10

A design proposal, not a tested system: a dual-view simulation loop in which an LLM-simulated user issues reflective instructions when about to disengage, and an instruction-following recommender adapts to maximize si...

Reference graph

Works this paper leans on

13 extracted references · 11 canonical work pages · cited by 1 Pith paper

[1]

Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. 2023. TWIN: TWo-stage Interest Network for Lifelong User Behavior Modeling in CTR Prediction at Kuaishou. In KDD

work page 2023
[2]

Zhichao Feng, JunJie Xie, Kaiyuan Li, Yu Qin, Pengfei Wang, Qianzhong Li, Bin Yin, Xiang Li, Wei Lin, and Shangguang Wang. 2024. Context-based Fast Recommendation Strategy for Long User Behavior Sequence in Meituan Waimai. In WWW. 355–363

work page 2024
[3]

Yinfeng Li, Chen Gao, Xiaoyi Du, Huazhou Wei, Hengliang Luo, Depeng Jin, and Yong Li. 2022. Automatically Discovering User Consumption Intents in Meituan. In KDD. 3259–3269

work page 2022
[4]

Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. In CIKM. 2685–2692

work page 2020
[5]

Yukun Ping, Chen Gao, Taichi Liu, Xiaoyi Du, Hengliang Luo, Depeng Jin, and Yong Li. 2021. User Consumption Intention Prediction in Meituan. In KDD. 3472–3482

work page 2021
[6]

Zihua Si, Lin Guan, ZhongXiang Sun, Xiaoxue Zang, Jing Lu, Yiqun Hui, Xingchao Cao, Zeyu Yang, Yichen Zheng, Dewei Leng, et al. 2024. Twin v2: Scaling ultra- long user behavior sequence modeling for enhanced ctr prediction at kuaishou. In CIKM. 4890–4897

work page 2024
[7]

Ying Sun, Yang Ji, Hengshu Zhu, Fuzhen Zhuang, Qing He, and Hui Xiong
[8]

Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. Dcn v2: Improved deep & cross network and practical lessons for web-scale learning to rank systems. In WWW. 1785–1797

work page 2021

Show all 13 references

[9]

Yuting Zhang, Yiqing Wu, Ran Le, Yongchun Zhu, Fuzhen Zhuang, Ruidong Han, Xiang Li, Wei Lin, Zhulin An, and Yongjun Xu. 2023. Modeling Dual Period-Varying Preferences for Takeaway Recommendation. In KDD

2023
[10]

Yongchun Zhu, Jingwu Chen, Ling Chen, Yitan Li, Feng Zhang, and Zuotao Liu
[11]

Yongchun Zhu, Zhenwei Tang, Yudan Liu, Fuzhen Zhuang, Ruobing Xie, Xu Zhang, Leyu Lin, and Qing He. 2022. Personalized transfer of user preferences for cross-domain recommendation. In WSDM. 1507–1515

2022
[12]

In SIGIR

Interest clock: Time perception in real-time streaming recommendation system. In SIGIR. 2915–2919
[2024]

ToIS (2024)

Market-aware Long-term Job Skill Recommendation with Explainable Deep Reinforcement Learning. ToIS (2024)

2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Jianxin Chang, Chenbin Zhang, Zhiyi Fu, Xiaoxue Zang, Lin Guan, Jing Lu, Yiqun Hui, Dewei Leng, Yanan Niu, Yang Song, et al. 2023. TWIN: TWo-stage Interest Network for Lifelong User Behavior Modeling in CTR Prediction at Kuaishou. In KDD

work page 2023

[2] [2]

Zhichao Feng, JunJie Xie, Kaiyuan Li, Yu Qin, Pengfei Wang, Qianzhong Li, Bin Yin, Xiang Li, Wei Lin, and Shangguang Wang. 2024. Context-based Fast Recommendation Strategy for Long User Behavior Sequence in Meituan Waimai. In WWW. 355–363

work page 2024

[3] [3]

Yinfeng Li, Chen Gao, Xiaoyi Du, Huazhou Wei, Hengliang Luo, Depeng Jin, and Yong Li. 2022. Automatically Discovering User Consumption Intents in Meituan. In KDD. 3259–3269

work page 2022

[4] [4]

Qi Pi, Guorui Zhou, Yujing Zhang, Zhe Wang, Lejian Ren, Ying Fan, Xiaoqiang Zhu, and Kun Gai. 2020. Search-based user interest modeling with lifelong sequential behavior data for click-through rate prediction. In CIKM. 2685–2692

work page 2020

[5] [5]

Yukun Ping, Chen Gao, Taichi Liu, Xiaoyi Du, Hengliang Luo, Depeng Jin, and Yong Li. 2021. User Consumption Intention Prediction in Meituan. In KDD. 3472–3482

work page 2021

[6] [6]

Zihua Si, Lin Guan, ZhongXiang Sun, Xiaoxue Zang, Jing Lu, Yiqun Hui, Xingchao Cao, Zeyu Yang, Yichen Zheng, Dewei Leng, et al. 2024. Twin v2: Scaling ultra- long user behavior sequence modeling for enhanced ctr prediction at kuaishou. In CIKM. 4890–4897

work page 2024

[7] [7]

Ying Sun, Yang Ji, Hengshu Zhu, Fuzhen Zhuang, Qing He, and Hui Xiong

[8] [8]

Ruoxi Wang, Rakesh Shivanna, Derek Cheng, Sagar Jain, Dong Lin, Lichan Hong, and Ed Chi. 2021. Dcn v2: Improved deep & cross network and practical lessons for web-scale learning to rank systems. In WWW. 1785–1797

work page 2021