Pith. sign in

REVIEW 3 major objections 3 minor 1 cited by

In multi-channel U2I2I retrieval, the paper shows that a trigger's value is the future engagement its retrieved items generate, and that this downstream-utility signal can be learned and coordinated across channels to improve recall and onl

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

CAPTS selects retrieval triggers by predicted downstream engagement per I2I channel and routes them jointly, reporting offline Recall@K gains and a small online time-spent lift.

T0 review reviewed 2026-08-02 challenge →

load-bearing objection Solid offline framework with a clean formulation, but the headline online lift is internally inconsistent and the offline metric is circular. the 3 major comments →

arxiv 2602.12564 v2 pith:3DXMKCY3 submitted 2026-02-13 cs.IR

CAPTS: Channel-Aware, Preference-Aligned Trigger Selection for Multi-Channel Item-to-Item Retrieval

Xiaoyou Zhou , Yuqi Liu , Zhao Liu , Xiao Lv , Bo Chen , Ruiming Tang , Guorui Zhou This is my paper
classification cs.IR
keywords Recommender SystemsTrigger SelectionItem-to-Item RetrievalMulti-Channel RetrievalValue AttributionRoutingDownstream UtilityShort-video Recommendation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper claims that in multi-channel item-to-item (U2I2I) recommendation, the right way to select trigger items is not by how much the trigger itself was liked or watched, but by the engagement that the items later retrieved from that trigger generate. It proposes CAPTS, a framework that learns channel-specific trigger scores from a look-ahead reward — summed watch time on retrieved items consumed in a future window — and routes triggers to channels under per-channel quotas while penalizing cross-channel overlap. Offline, CAPTS raises de-duplicated Recall@K across three I2I channels by up to 25% relative to the strongest rule-based baseline, and online A/B tests report a +0.351% lift in average time spent per device on a large short-video platform. The broader significance: if correct, trigger selection stops being a hand-tuned heuristic and becomes an end-to-end trained routing problem aligned with the actual objective of the retrieval stage.

Core claim

CAPTS defines a trigger's downstream utility V_c(t) as the expected sum of engagement over items that channel c would retrieve from trigger t and that the user consumes in a future window. The Value Attribution Module (VAM) turns this into per-trigger, per-channel supervision via timestamp-aligned I2I replay: for each logged request, it replays each channel seeded by each candidate trigger, intersects the retrieved set with the user's next w_s effective views, and sums watch time on matched items, yielding binary labels and intensity weights. The Channel-Adaptive Trigger Routing (CATR) module then learns channel-specific value predictions with a bounded calibrator and a uniqueness head that

What carries the argument

The central object is the forward-value definition of Eq. (2): V_c(t) = E[Σ_{j∈R_c(t,τ0)∩W_fut_u(τ0)} g(j)], which credits trigger t with the watch time on items that channel c retrieves from it and that the user actually consumes in the next w_s effective views. This definition redirects supervision from on-trigger feedback to downstream utility. VAM operationalizes it by replaying I2I retrieval at a timestamp-aligned snapshot, computing the intersection with the future window, and producing a clipped intensity label ℓ_c(t,τ0) and a binary label y_c(t,τ0) per channel. CATR consumes these labels to train channel-specific value heads with a calibrator and a uniqueness head ρ_c(t), and the rou

Load-bearing premise

The load-bearing premise is that watch time on an item falling in both the trigger's retrieved set and the future window was actually caused by that trigger, rather than by other triggers, the U2I channel, or item popularity; if this causality fails, VAM's rewards are biased and the learned routing can be misaligned.

What would settle it

For a given user and request, identify two triggers t1 and t2 that retrieve exactly the same set of items on a channel (same R_c), and compare the VAM rewards attributed to each under identical future consumption. If the rewards systematically differ because one trigger has higher on-trigger watch time or popularity — even though the retrieved item sets are identical — then the look-ahead signal is crediting item popularity or trigger-side engagement rather than downstream utility, contradicting the paper's causal interpretation.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Trigger selection ceases to be a per-channel, rule-based heuristic (Watch Time Top-N, Like Top-N) and becomes a learned routing task with downstream-utility supervision.
  • Coordinated routing with the uniqueness objective reduces cross-channel overlap in the merged candidate set, increasing diversity without sacrificing recall (Uniq@K rises across all channels).
  • The supervision pipeline requires only timestamp-aligned replay of each I2I channel, so the framework applies to any heterogeneous I2I channel set (collaborative filtering, model-based, multi-modal).
  • Because the retrieval stage sets the ceiling for the downstream ranking stage, the reported online lifts in DAU, total watch time, and average time per device indicate that improving trigger selection improves the entire cascade, not just retrieval metrics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The VAM reward is an attribution, not a causal estimate: it credits all future watch time on retrieved items to the trigger, so it may over-count items that would have been consumed anyway. A natural extension is to subtract a counterfactual baseline (e.g., engagement on the same items when recommended via U2I) to get a marginal trigger effect.
  • The uniqueness head is computed per trigger per channel; in a joint system, redundancy arises from the union of all triggers routed to a channel, so a set-level diversity term (e.g., avoiding duplicate items across all trigger seeds) could yield further gains.
  • With w_s=100 effective views as the default future window, CAPTS is tuned for short-horizon engagement. If the same framework were evaluated with longer horizons or temporal decay weighting, it might favour different triggers — e.g., novelty-seeking seeds — a testable variation the paper does not run.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes CAPTS, a unified trigger-selection framework for multi-channel U2I2I retrieval in industrial recommender systems. It introduces a Value Attribution Module (VAM) that labels triggers by the subsequent engagement on items retrieved from them, and a Channel-Adaptive Trigger Routing (CATR) module that learns channel-specific value and diversity-aware routing scores. Offline experiments on Kwai production data compare CAPTS to rule-based and learned baselines, reporting consistent Recall@K gains across three I2I channels. Online A/B tests report a +0.351% lift in average time spent per device. The central claim is that trigger selection can be learned end-to-end from downstream utility and coordinated across channels to reduce overlap and improve multi-channel recall.

Significance. The problem addressed is timely and practical: multi-channel U2I2I retrieval is central to large-scale recommender systems, and trigger selection is usually done by rules or per-channel heuristics. CAPTS offers a clean formulation—value attribution through logged replay and a routing objective with a diversity term—that is extensible and well-motivated. If the empirical claims hold, the framework could be a useful addition to the industrial toolkit. However, the evidence currently reported has a serious statistical inconsistency in the online lift and a circularity issue in the offline evaluation, both of which must be resolved before the claims can be accepted. The paper does not provide code or machine-checked proofs, and several citations are placeholders, which further limits verifiability.

major comments (3)
  1. [Section 4.4.1, Table 6] The reported online lift for average time spent per device is +0.351% with a 95% confidence interval of [-0.15%, +0.15%] and p<0.05. These numbers are mutually inconsistent: the interval excludes the point estimate, and a 95% CI containing zero implies p>0.05 for a two-sided test. At least one of the point estimate, the CI, or the significance claim is incorrect. Since this lift is cited in the Abstract and Conclusion as production evidence, the headline online claim is currently unsupported. Please correct the statistics or remove the claim.
  2. [Section 4.1.2 and Eq. (3)] The offline Recall@K metric is defined on the same future window W_fut_u(tau0) that VAM uses to construct supervision labels (Eq. 3 and Eq. 5). CAPTS is therefore trained to maximize the exact quantity on which it is evaluated, while the baselines receive no such look-ahead supervision. This makes the offline comparison circular: the gains in Tables 2 and 3 may reflect that the model is optimized for the evaluation target, not a genuine improvement in retrieval quality. To support the offline claim, evaluate on a held-out future window not used in training, or report alternative metrics (e.g., engagement-weighted recall with a fresh label horizon).
  3. [Eq. (3) and Appendix C] The VAM credit rule assumes that all future engagement on items in the intersection R_c(t,tau0) ∩ W_fut_u(tau0) is causally generated by trigger t on channel c. However, items can also be consumed through other triggers, U2I retrieval, or popularity. The paper does not provide a counterfactual or sensitivity analysis to validate this attribution, and the 20-minute replay rollback in Appendix C is asserted without supporting evidence. Without this, the supervision labels may be systematically biased. Please show that the attribution is robust, e.g., by comparing with a matched-control or leave-one-out trigger attribution.
minor comments (3)
  1. [References] Several citations are incomplete placeholders, e.g., 'Swing [? ]' in Sections 4.1.2 and 4.4.1 and in Appendix B. Please complete the reference list.
  2. [Section 4.1.2] The definition of Recall@K states that K is the per-channel cutoff, but later text refers to 'R@100' as a cutoff. Clarify whether K refers to the per-channel top-K or to the union size; Table 2 reports values for K=100..2000, which is consistent, but the wording should be tightened.
  3. [Figure 5 and Table 4] Figure 5 plots sensitivity to w_s for each channel, but the three panels are not explicitly labeled in the caption; also the table in Section 4.3.1 uses 'U@K' without defining it in the body. Add clear axis labels and define Uniq@K in the main text.

Circularity Check

2 steps flagged

Offline Recall@K and Uniq@K gains reduce to training on the evaluation metric itself; online lift is independent but statistically inconsistent.

specific steps
  1. fitted input called prediction [Section 3.1, Eqs. (2)-(5); Section 4.1.2 (Recall@K definition)]
    "V_c(t) ≜ E[ ∑_{j∈R_c(t,τ0)∩W_fut_u(τ0)} g(j) ] (Eq. 2); r_c(t,τ0) = ∑_{j∈R_c(t,τ0)∩W_fut_u(τ0)} g(j) (Eq. 3); ... Recall@K is defined as |W_fut_u(τ0) ∩ R_K_union| / |W_fut_u(τ0)|, i.e., the fraction of unique videos consumed in the future window that appear in R_K_union."

    The VAM supervision label y_c(t,τ0) used to train CAPTS is a thresholded version of the engagement-weighted intersection R_c(t,τ0)∩W_fut_u(τ0). The offline success metric Recall@K is the size of the same future-window intersection for the union of retrieved items, normalized by W_fut. CAPTS is therefore fit to maximize the exact quantity on which it is evaluated, while rule-based baselines never receive this future-window supervision. The offline 'prediction' is the training objective re-measured on held-out requests, not an independent test of retrieval quality.

  2. fitted input called prediction [Section 3.2, Eqs. (10)-(13); Section 4.3.1 (Uniq@K)]
    "The channel-unique subset is defined as: U_c(t,τ0) = { j ∈ R_c(t,τ0) : j ∉ R_c'(t,τ0) for all c'≠c } ... The uniqueness ratio is: ρ_c(t,τ0) = |U_c(t,τ0)| / |R_c(t,τ0)| + ε ... We compute Uniq_c@K = |U^K_c| / |R^K_c| + ε, which is consistent with Eq. (10) and Eq. (11)."

    The uniqueness supervision y_u_c(t,τ0) is formed by thresholding ρ_c(t,τ0) from Eq. (11), and the offline Uniq@K metric is the same ρ_c computed on Top-K results. The diversity head is trained to predict this exact definition and then evaluated with the identical formula. The reported reduction in cross-channel overlap is therefore the training target itself, not an independent diversity measurement.

full rationale

The offline central claim is partially circular. CAPTS's value labels are built from the future window W_fut (Eqs. 3-5), and offline Recall@K is defined on the same W_fut intersection; similarly, the uniqueness label and Uniq@K share one definition. Consequently, the offline gains over rule-based baselines are by construction aligned with the training objective, and the baselines are not given equivalent look-ahead supervision. This is not a complete definitional identity, so I score it 6 rather than 8-10. The online A/B test is genuinely out-of-sample and independent of the VAM training objective, but it contains an internal statistical inconsistency: the reported +0.351% lift is outside the stated 95% CI [-0.15%, +0.15%], and a 95% CI including zero is incompatible with the caption's p<0.05. That is a correctness problem, not a circularity. I also note no load-bearing self-citation: the overlapping-author citations (MARM, SIM, DIN, DIEN) are used as fixed retrieval channels or related work, not to justify CAPTS's central claim. The missing Swing citation and the untested 20-minute replay rollback in Appendix C are additional rigor concerns but do not affect the circularity score.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper introduces conceptual modules (VAM, CATR) rather than physical entities. The central empirical claim rests on a small set of hand-set hyperparameters and on the causal attribution assumption in Eq. (3), which is not independently validated.

free parameters (5)
  • Future window size w_s = 100 effective views
    Core VAM label horizon; selected after sensitivity sweep (Section 4.2.2); larger or smaller windows change labels and offline recall.
  • Channel scale s_c and cap M_c = s_c=100, M_c=6 (Appendix B)
    Hand-set channel-wise normalization and clipping for VAM intensity labels (Eq. 4), affecting all supervision.
  • Channel threshold gamma_c = not reported
    Binarizes VAM labels in Eq. (5); value is production-defined but not specified.
  • Routing/calibration hyperparameters = beta=0.1, theta=0.8, lambda=mu=0.1; eta not reported
    Trade-off weights in Eqs. (7), (14), (15); set by hand, no sensitivity analysis shown for most.
  • Replay rollback delta = ~20 minutes (Appendix C)
    Assumed lag between retrieval refresh and interaction stream; any mismatch biases VAM labels.
axioms (4)
  • domain assumption Engagement on items in R_c(t,t0) ∩ W_fut is causally generated by trigger t (Eq. 3).
    No counterfactual control; items could be consumed via U2I or other triggers/channels, and popular items will intersect many retrieved sets.
  • domain assumption Timestamp-aligned replay with 20-min rollback gives an unbiased estimate of what would have been retrieved at t0 (Appendix C).
    Assumes artifact snapshots and interaction availability are aligned enough; unverified if retrieval refresh lags differ.
  • domain assumption The multi-channel objective is additive over triggers and channels (Eq. 1) and per-channel budgets are fixed.
    Ignores item-level saturation and within-budget interactions except via the heuristic diversity term.
  • domain assumption Watch time on effective views is a sufficient engagement signal for preference alignment (Eq. 2).
    Other signals (likes, shares, follows) are present in logs but not used in V_c(t).

reviewed 2026-08-02 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CAPTS: Channel-Aware, Preference-Aligned Trigger Selection for Multi-Channel Item-to-Item Retrieval." pith.science (2026). https://pith.science/paper/3DXMKCY3

@misc{pith2026260212564,
  author       = {Pith},
  title        = {Pith review of: CAPTS: Channel-Aware, Preference-Aligned Trigger Selection for Multi-Channel Item-to-Item Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3DXMKCY3}},
  note         = {Machine review of arXiv:2602.12564}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large-scale industrial recommender systems commonly adopt multi-channel retrieval for candidate generation, combining direct user-to-item (U2I) retrieval with two-hop user-to-item-to-item (U2I2I) pipelines. In U2I2I, the system selects a small set of historical interactions as triggers to seed downstream item-to-item (I2I) retrieval across multiple channels. In production, triggers are often selected using rule-based policies or learned scorers and tuned in a channel-by-channel manner. However, these practices face two persistent challenges: biased value attribution that values triggers by on-trigger feedback rather than their downstream utility as retrieval seeds, and uncoordinated multi-channel routing where channels select triggers independently under a shared quota, increasing cross-channel overlap. To address these challenges, we propose Channel-Aware, Preference-Aligned Trigger Selection (CAPTS), a unified and flexible framework that treats multi-channel trigger selection as a learnable routing problem. CAPTS introduces a Value Attribution Module (VAM) that provides look-ahead supervision by crediting each trigger with the subsequent engagement generated by items retrieved from it on each I2I channel, and a Channel-Adaptive Trigger Routing (CATR) module that coordinates trigger-to-channel assignment to maximize the overall value of multi-channel retrieval. Extensive offline experiments and large-scale online A/B tests on Kwai, Kuaishou's international short-video platform, show that CAPTS consistently improves multi-channel recall offline and delivers a +0.351% lift in average time spent per device online.

Figures

Figures reproduced from arXiv: 2602.12564 by Bo Chen, Guorui Zhou, Ruiming Tang, Xiao Lv, Xiaoyou Zhou, Yuqi Liu, Zhao Liu.

Figure 1
Figure 1. Figure 1: Multi-channel retrieval with two candidate [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Challenges in multi-channel U2I2I trigger selec [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the CAPTS framework. (a) VAM: at re [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Production deployment of CAPTS. Orange lines de [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensitivity of Recall@K to the future window [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Channel-wise exposure share and engagement [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Multi-Decoder OneRec: Controllable Generative Retrieval for Multi-Objective Industrial Recommendation

    cs.IR 2026-07 conditional novelty 6.0

    Multi-Decoder OneRec attaches small LoRA experts for each objective to a shared generative recommender and uses quota-aware constrained beam search, beating single-decoder OneRec on offline recall and in production.

Reference graph

Works this paper leans on

30 extracted references · 5 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    Pith/arXiv arXiv 2025

  2. [2]

    Yue Cao, Xiaojiang Zhou, Jiaqi Feng, Peihao Huang, Yao Xiao, Dayao Chen, and Sheng Chen. 2022. Sampling Is All You Need on Modeling Long-Term User Behaviors for CTR Prediction. InCIKM. ACM, 2974–2983

    2022

  3. [3]

    Yukuo Cen, Jianwei Zhang, Xu Zou, Chang Zhou, Hongxia Yang, and Jie Tang

  4. [4]

    McAuley, and Caiming Xiong

    Yongjun Chen, Zhiwei Liu, Jia Li, Julian J. McAuley, and Caiming Xiong. 2022. Intent Contrastive Learning for Sequential Recommendation. InWWW. ACM, 2172–2182

    2022

  5. [5]

    Paul Covington, Jay Adams, and Emre Sargin. 2016. Deep Neural Networks for YouTube Recommendations. InRecSys. ACM, 191–198

    2016

  6. [6]

    JamesDavidson,BenjaminLiebald,JunningLiu,PalashNandy,TaylorVanVleet, UllasGargi,SujoyGupta,YuHe,MikeLambert,BlakeLivingston,andDasarathi Sampath. 2010. The YouTube video recommendation system. InProceedings of the Fourth ACM Conference on Recommender Systems (Barcelona, Spain)(RecSys ’10).AssociationforComputingMachinery,NewYork,NY,USA,293–296. doi:10. 11...

    arXiv 2010

  7. [7]

    ChantatEksombatchai,PranavJindal,JerryZitaoLiu,YuchenLiu,RahulSharma, Charles Sugnet, Mark Ulrich, and Jure Leskovec. 2018. Pixie: A System for Rec- ommending3+BillionItemsto200+MillionUsersinReal-Time.In WWW.ACM, 1775–1784

    2018

  8. [8]

    Ahmed El-Kishky, Thomas Markovich, Serim Park, Chetan Verma, Baekjin Kim, Ramy Eskander, Yury Malkov, Frank Portman, Sofía Samaniego, Ying Xiao, and Aria Haghighi. 2022. TwHIN: Embedding the Twitter Heterogeneous Informa- tion Network for Personalized Recommendation. InKDD. ACM, 2842–2850

    2022

  9. [9]

    Gomez-Uribe and Neil Hunt

    Carlos A. Gomez-Uribe and Neil Hunt. 2016. The Netflix Recommender System: Algorithms, Business Value, and Innovation.ACM Trans. Manage. Inf. Syst. 6, 4, Article 13 (Dec. 2016), 19 pages.doi:10.1145/2843948

    doi:10.1145/2843948 2016

  10. [10]

    Po-Sen Huang, Xiaodong He, Jianfeng Gao, Li Deng, Alex Acero, and Larry P. Heck. 2013. Learning deep structured semantic models for web search using clickthrough data. InCIKM. ACM, 2333–2338

    2013

  11. [11]

    Liu, Eric Tzeng, and Yushi Jing

    Dmitry Kislyuk, Yuchen Liu, David C. Liu, Eric Tzeng, and Yushi Jing. 2015. Hu- man Curation and Convnets: Powering Item-to-Item Recommendations on Pin- terest. CoRR abs/1511.04003 (2015)

    Pith/arXiv arXiv 2015

  12. [12]

    Chao Li, Zhiyuan Liu, Mengmeng Wu, Yuchi Xu, Huan Zhao, Pipei Huang, Guo- liang Kang, Qiwei Chen, Wei Li, and Dik Lun Lee. 2019. Multi-Interest Network with Dynamic Routing for Recommendation at Tmall. InCIKM. ACM, 2615– 2623

    2019

  13. [13]

    Path-basedDeepNetworkforCandidate Item Matching in Recommenders

    Houyi Li, Zhihong Chen, Chenliang Li, Rong Xiao, Hongbo Deng, Peng Zhang, YongchaoLiu,andHaihongTang.2021. Path-basedDeepNetworkforCandidate Item Matching in Recommenders. InSIGIR. ACM, 1493–1502

    2021

  14. [14]

    MARM:UnlockingtheFu- ture of Recommendation Systems through Memory Augmentation and Scalable Complexity

    Xiao Lv, Jiangxia Cao, Shijie Guan, Xiaoyou Zhou, Zhiguang Qi, Yaqiang Zang, MingLi,BenWang,KunGai,andGuoruiZhou.2024. MARM:UnlockingtheFu- ture of Recommendation Systems through Memory Augmentation and Scalable Complexity. CoRR abs/2411.09425 (2024)

    Pith/arXiv arXiv 2024

  15. [15]

    Yue Meng, Cheng Guo, Xiaohui Hu, Honghu Deng, Yi Cao, Tong Liu, and Bo Zheng. 2025. User Long-Term Multi-Interest Retrieval Model for Recommenda- tion. InRecSys. ACM, 1112–1116

    2025

  16. [16]

    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. InCIKM. ACM, 2685–2692

    2020

  17. [17]

    Aneesh Sharma, Jerry Jiang, Praveen Bommannavar, Brian Larson, and Jimmy Lin. 2016. GraphJet: Real-Time Content Recommendations at Twitter. Proc. VLDB Endow. 9, 13 (2016), 1281–1292

    2016

  18. [18]

    JizheWang,PipeiHuang,HuanZhao,ZhiboZhang,BinqiangZhao,andDikLun Lee. 2018. Billion-scale Commodity Embedding for E-commerce Recommenda- tion in Alibaba. InKDD. ACM, 839–848

    2018

  19. [19]

    XuXie,FeiSun,ZhaoyangLiu,ShiwenWu,JinyangGao,JiandongZhang,Bolin Ding, and Bin Cui. 2022. Contrastive Learning for Sequential Recommendation. In ICDE. IEEE, 1259–1273

    2022

  20. [20]

    Jing Yan, Liu Jiang, Jianfei Cui, Zhichen Zhao, Xingyan Bin, Feng Zhang, and Zuotao Liu. 2024. Trinity: Syncretizing Multi-/Long-Tail/Long-Term Interests All in One. InKDD. ACM, 6095–6104

    2024

  21. [21]

    Xiaoyong Yang, Yadong Zhu, Yi Zhang, Xiaobo Wang, and Quan Yuan. 2020. Large Scale Product Graph Construction for Recommendation in E-commerce. CoRR abs/2010.05525 (2020)

    Pith/arXiv arXiv 2020

  22. [22]

    Hamilton, and Jure Leskovec

    Rex Ying, Ruining He, Kaifeng Chen, Pong Eksombatchai, William L. Hamilton, and Jure Leskovec. 2018. Graph Convolutional Neural Networks for Web-Scale Recommender Systems. InKDD. ACM, 974–983

    2018

  23. [23]

    Chi Zhang, Yantong Du, Xiangyu Zhao, Qilong Han, Rui Chen, and Li Li. 2022. Hierarchical Item Inconsistency Signal Learning for Sequence Denoising in Se- quential Recommendation. InCIKM. ACM, 2508–2518

    2022

  24. [24]

    Jiang Zhang, Sumit Kumar, Wei Chang, Yubo Wang, Feng Zhang, Weize Mao, Hanchao Yu, Aashu Singh, Min Li, and Qifan Wang. 2025. Optimizing Recall or Relevance? A Multi-Task Multi-Head Approach for Item-to-Item Retrieval in Recommendation. CoRR abs/2506.06239 (2025)

    Pith/arXiv arXiv 2025

  25. [25]

    Jianjin Zhang, Zheng Liu, Weihao Han, Shitao Xiao, Ruicheng Zheng, Yingxia Shao, Hao Sun, Hanqing Zhu, Premkumar Srinivasan, Weiwei Deng, Qi Zhang, and Xing Xie. 2022. Uni-Retriever: Towards Learning the Unified Embedding Based Retriever in Bing Sponsored Search. InKDD. ACM, 4493–4501

    2022

  26. [26]

    Guorui Zhou, Na Mou, Ying Fan, Qi Pi, Weijie Bian, Chang Zhou, Xiaoqiang Zhu, and Kun Gai. 2019. Deep Interest Evolution Network for Click-Through Rate Prediction. InAAAI. AAAI Press, 5941–5948

    2019

  27. [27]

    Guorui Zhou, Xiaoqiang Zhu, Chengru Song, Ying Fan, Han Zhu, Xiao Ma, Yanghui Yan, Junqi Jin, Han Li, and Kun Gai. 2018. Deep Interest Network for Click-Through Rate Prediction. InKDD. ACM, 1059–1068

    2018

  28. [28]

    YongchunZhu,GuanyuJiang,JingwuChen,FengZhang,QiWu,andZuotaoLiu

  29. [2020]

    Controllable Multi-Interest Framework for Recommendation. In KDD. ACM, 2942–2951

  30. [2025]

    InWWW (Companion Volume)

    Long-Term Interest Clock: Fine-Grained Time Perception in Streaming Recommendation System. InWWW (Companion Volume). ACM, 1554–1557. A Notation Table7 summarizes the main notation. B Implementation Details In offline supervision construction, we perform I2I replay on can- didatetriggersforeachrequesttoobtainchannel-specificretrieval resultsandbuildVAMsupe...

    2017

This paper was first reviewed by deepseek-v4-flash on August 2, 2026.