Pith. sign in

REVIEW 2 major objections 5 minor 33 references

TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising

T0 review · 2 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that long-horizon conversion prediction improves when target-window conversion probability and conversion-delay distribution are learned separately from two event clocks, yielding monotone horizon predictions and a 2.5% pro

desk verdict Genuinely new two-clock delayed-feedback CVR with strong production evidence; the self-referential exposure-mass loop needs an in-situ calibration check before I'd fully trust it. read the letter →

arxiv 2607.25404 v1 pith:BJ7KC2DF submitted 2026-07-28 cs.LG cs.IR

Kaiyuan Li , Kun Wang , Zhongbo Wang , Teng Sha , Ming Yan , Yanhua Cheng , Xialong Liu This is my paper · ORCID
classification cs.LGcs.IR
keywords delayedfeedbackconversionratepredictiononlineadvertisingtwo-clocklearningtemporalconvolutioncensoredlabelslong-horizonincrementaltraining
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

TWICE is an attempt to solve a two-clock learning problem in online advertising: clicks are released for training after a short observation window, while conversions keep arriving over a much longer target window. The paper's load-bearing idea is to factor the long-horizon conversion probability into a target-window CVR head and a grouped elapsed-delay CDF, and to train each factor from the clock that observes it best—fresh clicks supervise CVR through a short-window status likelihood with a stopped delay factor, while late-arriving conversions supervise only the delay head through an arrival-conditioned likelihood that weights each historical cohort by its frozen predicted conversion mass. If the factorization is right, one checkpoint yields monotone predictions for every horizon up to the target window, and a production A/B test shows expected revenue up 2.486%, conversions up 2.061%, and calibration (PCOC) improving from 0.961 to 0.990, all at p<0.01.

What carries the argument

The central machinery is the two-clock/two-window factorization p̂_u(x) = p_θ(x)·F_φ(u|G(x)) combined with an arrival-conditioned temporal-convolution likelihood. The delay head is a softmax over K elapsed-delay buckets whose prefix sums form a monotone CDF; at each conversion arrival, timestamp range sums over the logged pCVR index give R_{i,k}, and the likelihood compares the model's arrival intensity Λ_i to the realized bucket. Stop-gradients keep click-clock and conversion-clock supervision on disjoint parameter sets, so late conversions improve timing estimates without replaying historical clicks through the CVR backbone, and one checkpoint serves every horizon u≤v by multiplying p_θ by

What would settle it

Take mature 30-day data, split conversions by realized delay bucket and delay context G(x), and compare the average logged pCVR mass of converters in each bucket with that of non-converters; if the two diverge (calibration error correlates with delay), the arrival-conditioned likelihood is biased and the learned F_φ will not reproduce the empirical within-group delay CDF—a per-group KS test between F_φ and the empirical CDF would detect the failure.

Watch

Extended reading notes

Core claim

On a public delayed-feedback benchmark and a large industrial dataset, the paper claims that the two-clock split recovers 82–90% of the recoverable gap between training on one-hour-old clicks and training on mature 30-day labels, beating all deployable baselines. The central formula is p̂_u(x) = p_θ(x)·F_φ(u|G(x)): the CVR head p_θ estimates conversion within the target window v, the delay head F_φ estimates the fraction of those converters that have arrived by horizon u, and G(x) is a compact grouping of clicks with similar delay behavior. TWICE trains p_θ on click-clock records using the base-window status Y^(o) with the delay factor frozen (stop-gradient), and trains F_φ on conversion-clo

Load-bearing premise

The delay likelihood is exact only if the model's own logged predicted conversion rates are calibrated in aggregate within each delay context and click-time range and delay distributions are stationary there; since that mass is the model's self-generated prediction and the CVR loss sees only the product, miscalibration correlated with conversion delay would bias the delay head and silently leak into every horizon prediction.

Editorial extensions

If this is right

  • One trained TWICE checkpoint produces monotone CVR estimates for every horizon up to the 30-day target window; horizons shorter than the base observation window are handled by the same model without retraining.
  • Because late conversions update only a compact delay head via scalar pCVR masses, the delay stream does not require historical click features, keeping hourly training and 36 ms serving latency unchanged.
  • On both the public and industrial datasets, TWICE recovers roughly 82–90% of the Vanilla-to-Oracle headroom, and the relative gain over the strongest baseline grows with conversion delay—from a 0.00012 AUC gain for sub-hour delays to 0.00405 for 14–30 day delays on the public benchmark.
  • In the production A/B test, expected revenue, revenue, and conversions rose 2.486%, 1.858%, and 2.061% (all p<0.01) while PCOC moved from 0.961 to 0.990, and the system was subsequently deployed to full traffic.

Reading between the lines

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

  • Because the delay head is trained with the model's own stopped pCVR predictions as exposure, the method's unbiasedness rests on p_θ being calibrated within each delay context and click-time range; an in-situ calibration check of logged a_j versus mature outcomes, stratified by G(x) and click time, would make the assumption testable and could be added as a production monitoring signal.
  • The paper's working approximation F(b_m|x)≈F(b_m|G(x)) means horizon-monotonicity holds per group, not per click; a natural extension is a hierarchical delay context that gives finer per-click delay estimates while falling back to a coarser key to keep risk sets non-singleton, which the paper itself identifies as the reason it avoids full-feature keys.
  • The production A/B measures full-system lift, not the mechanism in isolation; an online or large-scale offline ablation that disables only the pCVR-weighting in the arrival-conditioned likelihood would test whether cohort-exposure correction, rather than the factorization alone, drives the gain.
  • If delay distributions drift within a group due to policy or product changes, the stationarity assumption of Section 4.3 is violated; conditioning the delay head on calendar time or allowing a slowly varying delay kernel is a testable extension for non-stationary regimes.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. The paper proposes TWICE, a delayed-feedback CVR model that factorizes the target-window conversion probability into a full-feature CVR head p_theta(x) and a grouped, elapsed-delay CDF F_phi(u|G(x)). The CVR head is trained on fresh click-clock records with a stopped delay factor (Eq. 9–10), while the delay head is trained on conversion-clock arrivals through an arrival-conditioned temporal-convolution objective that weights historical cohorts by stopped pCVR mass (Eq. 17–21). The authors claim monotone multi-horizon predictions from a single checkpoint, no serving-time historical lookup, and report offline gains on Criteo and an industrial dataset, plus an online A/B test with statistically significant lifts in expected revenue, revenue, and conversions.

Significance. If the claims hold, TWICE is a practically valuable contribution: it provides a clean two-clock formulation of delayed feedback, a monotone CDF for arbitrary horizons up to the target window, and an efficient training scheme requiring only a scalar pCVR-mass index rather than historical feature replay. The paper is unusually transparent: Section 4.3 states the exactness conditions for Eq. (21), and Appendices C–D explicitly disclaim statistical consistency, proving only mass conservation and gradient equivalence for fixed stopped scores. The ablations (Table 3) and the delayed-feedback-stratified analysis (Figure 5) are informative and support the claim that the temporal-convolution mechanism, not just the factorization, drives the gains. The main weakness is that a central assumption—aggregate calibration of the self-generated pCVR mass within delay contexts and click-time ranges—is stated but never directly verified in situ, leaving a plausible feedback-loop bias in the very quantity the method improves.

major comments (2)
  1. [§4.3, Eq. (21)] The arrival-conditioned likelihood is exact only when the logged pCVR mass a_j = sg[p_theta(x_j)] is calibrated in aggregate within each delay context and click-time range, and when within-group delay is stationary. Since a_j is the model's own prediction, and since the click-clock loss (Eq. 10) supervises only the product p_theta * sg[F_phi(o|g)], any miscalibration of a_j that correlates with conversion delay biases F_phi, and that bias is then absorbed into p_theta through the stopped delay factor. The paper acknowledges these conditions are only approximated and explicitly disclaims statistical consistency, but it reports no in-situ calibration diagnostic for a_j (e.g., reliability by group and click-time cohort, or a sensitivity analysis with recalibrated a_j). Table 3's w/o pCVR Weighting ablation is a useful stress test but does not substitute for verifying the calibration conditi
  2. [§4.2, Eq. (11) discussion] The working approximation F(b_m|x) ≈ F(b_m|G(x)) is load-bearing for the design of the delay head, but the paper does not provide a direct measure of how much delay heterogeneity is lost by conditioning on a compact group G(x). Table 4 shows that doubling a 10-field context to a 20-field context yields little gain, which is reassuring, but it does not quantify the approximation error relative to the full-feature delay CDF. Given that the full-x objective is degenerate by construction, this may be the best feasible check, but a short analysis of within-group delay heterogeneity (e.g., group-level delay CDF variance or a calibration plot of F vs. group) would strengthen the claim that the grouping sufficiency is not a major source of bias.
minor comments (5)
  1. [§5.7] The A/B test baseline is described only as 'an enhanced FSIW-style baseline sharing the same features, conversion definition, deployed backbone, and serving stack.' Please specify what enhancements are applied and whether the comparison is apples-to-apples in training data and model capacity. Also report confidence intervals for the lifts, not just p-values.
  2. [Table 4] The 'Fields' column formatting is difficult to read (e.g., 'Minimal 𝐺 0 6'). Please align the numbers and clarify whether 0 fields for Global means a single shared context key.
  3. [Figure 2] The annotations 'stop gradient ×', 'training only', and 'only' are dense and partly illegible in the reproduction. Consider separating the gradient-path diagram from the data-flow diagram for readability.
  4. [Appendix E] lambda_d = 1 is set without sensitivity analysis. Since Eq. (22) jointly trains two loss terms, a brief lambda_d sweep or a statement that results are stable would help.
  5. [§5.6] The traffic-dynamics case study is qualitative ('visually closer'). Consider adding a quantitative agreement measure (e.g., correlation or Wasserstein distance) for the reconstruction in Figure 6(b).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: TWICE's factorization and delay objective are plug-in modeling choices, not predictions derived from their own inputs.

full rationale

The derivation chain is self-contained and does not reduce to its inputs by construction. Equation 6 is a tautological decomposition p_o(x)=p_v(x)F(o|x), and Equation 15 defines the horizon predictor as a product of two separately parameterized heads; neither equation defines a fitted quantity as the predicted quantity. The click-clock loss (Eqs. 9–10) supervises the product p_theta * sg[F_phi(o)] against observed short-window labels, while the delay loss (Eqs. 17–21) uses stopped pCVR mass a_j = sg[p_theta(x_j)] as an exposure weight. That is a plug-in/EM-style estimation loop, not a definitional circle: F_phi is not defined as p_o/p_theta, and p_theta is not defined as p_o/F_phi. The paper itself flags the key caveat in Section 4.3 ('Assumptions and Scope'): Equation 21 is exact only when logged pCVR mass is calibrated in aggregate within delay context and click-time range and delays are stationary; it explicitly disclaims universally unbiased delay estimation, and Appendices C and D prove only algebraic identities for fixed stopped scores, not statistical consistency. This is a calibration/robustness risk, not a circularity. There is no load-bearing self-citation: DFM and PI are external prior work, and the empirical claims are supported by public-benchmark comparisons and an online A/B test. The self-referential mechanism flagged by the reader is a legitimate identifiability concern but does not make the central claim equivalent to its inputs.

Assumptions & free parameters 6 free parameters · 6 assumptions · 2 invented entities

TWICE's contribution is a training regimen, not a set of new physical postulates: the only 'entities' are a conceptual framing and a data structure. The free parameters are hyperparameters of the streaming design (o, bucket partition, group rule, λ_d) plus the self-generated pCVR masses that couple the two heads. The axioms are the standard DFM factorization, the paper's own stated exactness conditions for Eq. 21 (calibration and stationarity, §4.3), the superposition assumption behind Eq. 7, and two explicitly admitted approximations (grouping sufficiency and piecewise interpolation). The load-bearing count is axioms 2 and 3: if pCVR mass is miscalibrated or delay non-stationary, the delay CDF and hence all horizon predictions shift.

free parameters (6)
  • λ_d (delay loss weight) = 1
    Chosen by hand in Appendix E ('TWICE uses o = 1 hour and λ_d = 1'); no sensitivity analysis is reported, so the balance of the joint objective is unexamined.
  • Base observation window o = 1 hour
    Design choice (Appendix E); §5.4 tests o ∈ {1,2,4,8,12,24h} and shows robustness, but the deployed value is arbitrary within that range.
  • Delay bucket partition B_1..B_12 = 0,1,2,4,8,12,24h,2,4,7,14,21,30d
    Hand-chosen discretization of [0,30d] (Appendix E); the piecewise-uniform interpolation and softmax PMF depend on these boundaries.
  • Delay context G(x) = Default: 10 stable fields (industrial); all released categoricals (Criteo)
    Hand-chosen trade-off between timing heterogeneity and risk-set size (§4.3, §5.3.2); expanded context gives only +0.0002 AUC, so the choice is empirically bracketed but not principled.
  • Stopped pCVR mass a_j = sg[p_θ(x_j)] = CVR-head outputs at click time
    Used as exposure weights in Eq. 18; they are fitted model outputs, so the delay head's target depends on the CVR head's calibration — the self-referential coupling identified in §4.3.
  • Learning rates = selected per method from 7 values
    Picked on a held-out chronological pretraining slice (§5.1); standard hyperparameter selection, not target-fitting.
assumptions (6)
  • domain assumption Delay-CDF factorization p_o(x) = p_v(x)·F(o|x) (Eq. 6)
    The short-window CVR is the long-window CVR times the fraction of converters observed by age o; taken from DFM, requires the label identity Y^(o) = Y^(v)·1(D≤o) to model cleanly.
  • domain assumption pCVR mass calibration: a_j ≈ p_v(x_j) in aggregate within each delay context and click-time range
    Stated in §4.3 'Assumptions and Scope' as an exactness condition for Eq. 21; it makes the delay model's target equal the true delay distribution. Never verified at group level.
  • domain assumption Stationary within-group delay distribution over the retained target window
    §4.3: 'within-context converters share a stationary delay distribution over the retained target window'; the paper admits production only approximates this.
  • domain assumption Superposition/convolvability of arrival intensity: Λ(t) = (m*f)(t) (Eq. 7)
    Conversion arrivals from distinct historical cohorts add linearly; Appendix D's gradient-equivalence argument relies on this linear structure.
  • ad hoc to paper Grouping sufficiency: F(b_m|x) ≈ F(b_m|G(x)) for m=1..K (Eq. 11 discussion)
    Full-feature conditioning would make risk sets singleton and the delay loss gradient-free (§4.3 'Why not a Full-x Aggregation Key?'); the compact G is a working approximation.
  • ad hoc to paper Piecewise-uniform delay CDF inside buckets (Eq. 14)
    Interpolation used to answer horizon queries u inside a bucket; convenient and monotone but not derived from a delay model.
invented entities (2)
  • Two-clock abstraction (click clock vs. conversion clock)
    purpose: Organizing frame for decomposing supervision into a fresh-status CVR loss and an arrival-conditioned delay loss
    A conceptual lens, not a testable physical or data entity; its utility is demonstrated only through the authors' own ablations.
  • Stopped pCVR mass index (a_j records keyed by click time and group)
    purpose: Provides cohort exposure weights R_{i,k} in the arrival-conditioned delay likelihood (Eqs. 17–18) without replaying historical features
    An engineering construct; supported by the paper's internal ablations (w/o-pCVR-weighting) but no external falsification handle independent of the authors' experiments.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising." pith.science (2026). https://pith.science/paper/BJ7KC2DF

@misc{pith2026260725404,
  author       = {Pith},
  title        = {Pith review of: TWICE: Two-Clock, Two-Window Learning for Long-Horizon Conversion Prediction in Online Advertising},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BJ7KC2DF}},
  note         = {Machine review of arXiv:2607.25404}
}
read the original abstract

Long-horizon conversion prediction under delayed feedback creates a two-clock, two-window learning problem in online advertising. A short base observation window releases recent clicks on the click clock before their outcomes mature, whereas conversions continue to arrive on the conversion clock throughout a longer target conversion window. The click clock provides timely but partially observed status supervision. The conversion clock reveals long-tail delays, but the delay composition within an arrival-time slice is weighted by historical click cohorts with different traffic volumes and target-window conversion rates. We present TWICE, a framework that factorizes long-horizon post-click conversion rate (CVR) into a target-window conversion probability and a grouped elapsed-delay cumulative distribution function (CDF). The two clocks provide complementary supervision. Click-clock records train the target-window CVR head through a current-status likelihood over the base observation window. Newly arrived conversions train the delay model on the conversion clock. To account for the cohort mixture, TWICE uses fixed click-time predicted CVR (pCVR) mass as cohort exposure in an arrival-conditioned likelihood. This accounts for differences in cohort traffic and conversion propensity. The resulting aggregate records are self-contained. A single learned CDF produces monotone predictions for all requested horizons up to the target conversion window. Serving requires neither historical lookup nor convolution. Experiments on a public benchmark and an industrial advertising dataset demonstrate the effectiveness of TWICE. In an online A/B test in Kwai's advertising system, TWICE increased expected revenue, revenue, and conversions by 2.486%, 1.858%, and 2.061%, respectively. It was subsequently deployed to full traffic.

Figures

Figures reproduced from arXiv: 2607.25404 by the authors.

Figure 1
Figure 1. Two-clock, two-window view of incremental de [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of TWICE. (a) Click-clock records train the CVR tower (fresh-status loss) and conversion-clock records train [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. TWICE’s hourly loop. Clicks log stopped pCVR [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Window scaling on Criteo, using setting-matched [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: Conversion-delay-stratified performance on Criteo. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 4 canonical work pages

  1. [1]

    Ashwinkumar Badanidiyuru, Andrew Evdokimov, Vinodh Krishnan, Pan Li, Wynn Vonnegut, and Jayden Wang. 2021. Handling Many Conversions per Click in Modeling Delayed Feedback. arXiv:2101.02284 [cs.LG] https://arxiv.org/abs/ 2101.02284

    arXiv 2021

  2. [2]

    Ashwinkumar Badanidiyuru Varadaraja, Zhe Feng, Tianxi Li, and Haifeng Xu

  3. [3]

    Olivier Chapelle. 2014. Modeling Delayed Feedback in Display Advertising. In Proceedings of the 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’14). Association for Computing Machinery, New York, NY, USA, 1097–1105. doi:10.1145/2623330.2623634

    arXiv 2014

  4. [4]

    Yu Chen, Jiaqi Jin, Hui Zhao, Pengjie Wang, Guojun Liu, Jian Xu, and Bo Zheng

  5. [5]

    Sunhao Dai, Yuqi Zhou, Jun Xu, and Ji-Rong Wen. 2023. Dually Enhanced Delayed Feedback Modeling for Streaming Conversion Rate Prediction. InPro- ceedings of the 32nd ACM International Conference on Information and Knowledge Management (CIKM ’23). Association for Computing Machinery, New York, NY, USA, 390–399. doi:10.1145/3583780.3614856

    arXiv 2023

  6. [6]

    InProceedings of the ACM Web Conference 2022 (WWW ’22)

    Asymptotically Unbiased Estimation for Delayed Feedback Modeling via Label Correction. InProceedings of the ACM Web Conference 2022 (WWW ’22). Association for Computing Machinery, New York, NY, USA, 369–379. doi:10. 1145/3485447.3511965

    arXiv 2022

  7. [7]

    Hui Gao and Yihan Yang. 2022. Multi-Head Online Learning for Delayed Feedback Modeling. arXiv:2205.12406 [cs.LG] https://arxiv.org/abs/2205.12406

    arXiv 2022

  8. [8]

    Chenlu Ding, Jiancan Wu, Yancheng Yuan, Cunchun Li, Xiang Wang, Dingxian Wang, Frank Yang, and Andrew Rabinovich. 2026. Delayed Feedback Model- ing with Influence Functions.Proceedings of the AAAI Conference on Artificial Intelligence40, 17 (2026), 14648–14656. doi:10.1609/aaai.v40i17.38483

    work page doi:10.1609/aaai.v40i17.38483 2026

Show all 33 references
  1. [9]

    Zhigang Huangfu, Gongduo Zhang, Zhengwei Wu, Qintong Wu, Zhiqiang Zhang, Lihong Gu, Jun Zhou, and Jinjie Gu. 2022. A Multi-Task Learning Approach for Delayed Feedback Modeling. InCompanion Proceedings of the Web Conference 2022 (WWW ’22 Companion). Association for Computing Ma...

    2022

  2. [10]

    Siyu Gu, Xiang-Rong Sheng, Ying Fan, Guorui Zhou, and Xiaoqiang Zhu. 2021. Real Negatives Matter: Continuous Training with Real Negatives for Delayed Feedback Modeling. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining (KDD ’21). Association ...

    2021

  3. [11]

    Kuang-chih Lee, Burkay Orten, Ali Dasdan, and Wentong Li. 2012. Estimating Conversion Rate in Display Advertising from Past Performance Data. InProceed- ings of the 18th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’12). Association for Compu...

    2012

  4. [12]

    Sofia Ira Ktena, Alykhan Tejani, Lucas Theis, Pranay Kumar Myana, Deepak Dilipkumar, Ferenc Huszar, Steven Yoo, and Wenzhe Shi. 2019. Addressing Delayed Feedback for Continuous Training with Neural Networks in CTR Pre- diction. InProceedings of the 13th ACM Conference on Recom...

    2019

  5. [13]

    Qiming Liu, Xiang Ao, Yuyao Guo, and Qing He. 2024. Online Conversion Rate Prediction via Multi-Interval Screening and Synthesizing under Delayed Feedback.Proceedings of the AAAI Conference on Artificial Intelligence38, 8 (2024), 8796–8804. doi:10.1609/aaai.v38i8.28726

    2024 doi

  6. [14]

    Qiming Liu, Haoming Li, Xiang Ao, Yuyao Guo, Zhihong Dong, Ruobing Zhang, Qiong Chen, Jianfeng Tong, and Qing He. 2023. Online Conversion Rate Predic- tion via Neural Satellite Networks in Delayed Feedback Advertising. InProceed- ings of the 46th International ACM SIGIR Confer...

    2023

  7. [15]

    Haoming Li, Feiyang Pan, Xiang Ao, Zhao Yang, Min Lu, Junwei Pan, Dapeng Liu, Lei Xiao, and Qing He. 2021. Follow the Prophet: Accurate Online Conversion Rate Prediction in the Face of Delayed Feedback. InProceedings of the 44th International ACM SIGIR Conference on Research a...

    2021

  8. [16]

    Xiao Ma, Liqin Zhao, Guan Huang, Zhi Wang, Zelin Hu, Xiaoqiang Zhu, and Kun Gai. 2018. Entire Space Multi-Task Model: An Effective Approach for Estimating Post-Click Conversion Rate. InThe 41st International ACM SIGIR Conference on Research & Development in Information Retriev...

    2018 doi

  9. [17]

    Yuta Saito, Gota Morishita, and Shota Yasui. 2020. Dual Learning Algorithm for Delayed Conversions. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’20). Association for Computing Machinery, New York, NY,...

    2020

  10. [18]

    Mingxuan Luo, Guipeng Xv, Sishuo Chen, Xinyu Li, Li Zhang, Zhangming Chan, Xiang-Rong Sheng, Han Zhu, Jian Xu, Bo Zheng, and Chen Lin. 2026. Modeling Cascaded Delay Feedback for Online Net Conversion Rate Prediction: Benchmark, Insights and Solutions. InProceedings of the ACM ...

    2026

  11. [19]

    Yumin Su, Liang Zhang, Quanyu Dai, Bo Zhang, Jinyao Yan, Dan Wang, Yongjun Bao, Sulong Xu, Yang He, and Weipeng Yan. 2020. An Attention-based Model for Conversion Rate Prediction with Delayed Feedback via Post-click Calibration. InProceedings of the Twenty-Ninth International ...

    2020 doi

  12. [20]

    Hao Wang, Tai-Wei Chang, Tianqiao Liu, Jianmin Huang, Zhichao Chen, Chao Yu, Ruopeng Li, and Wei Chu. 2022. ESCM2: Entire Space Counterfactual Multi- Task Model for Post-Click Conversion Rate Estimation. InProceedings of the 45th International ACM SIGIR Conference on Research ...

    2022

  13. [21]

    Hongzu Su, Lichao Meng, Lei Zhu, Ke Lu, and Jingjing Li. 2024. DDPO: Direct Dual Propensity Optimization for Post-Click Conversion Rate Estimation. In Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’24). A...

    2024

  14. [22]

    Hong Wen, Jing Zhang, Yuan Wang, Fuyu Lv, Wentian Bao, Quan Lin, and Keping Yang. 2020. Entire Space Multi-Task Modeling via Post-Click Behavior Decom- position for Conversion Rate Prediction. InProceedings of the 43rd International ACM SIGIR Conference on Research and Develop...

    2020

  15. [23]

    Jiaqi Yang and De-Chuan Zhan. 2022. Generalized Delayed Feedback Model with Post-Click Information in Recommender Systems. InAdvances in Neural Information Processing Systems, Vol. 35. Curran Associates, Inc., New Orleans, LA, USA, 26192–26203. https://proceedings.neurips.cc/p...

    2022

  16. [24]

    Yifan Wang, Peijie Sun, Min Zhang, Qinglin Jia, Jingjie Li, and Shaoping Ma. 2023. Unbiased Delayed Feedback Label Correction for Conversion Rate Prediction. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’23). Association for Compu...

    2023

  17. [25]

    Shota Yasui and Masahiro Kato. 2022. Learning Classifiers under Delayed Feed- back with a Time Window Assumption. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’22). Association for Computing Machinery, New York, NY, USA, 2286–2295...

    2022 doi

  18. [26]

    Shota Yasui, Gota Morishita, Komei Fujita, and Masashi Shibata. 2020. A Feedback Shift Correction in Predicting Conversion Rates under Delayed Feedback. In Proceedings of The Web Conference 2020 (WWW ’20). Association for Computing Machinery, New York, NY, USA, 2740–2746. doi:...

    2020

  19. [27]

    Jia-Qi Yang, Xiang Li, Shuguang Han, Tao Zhuang, De-Chuan Zhan, Xiaoyi Zeng, and Bin Tong. 2021. Capturing Delayed Feedback in Conversion Rate Prediction via Elapsed-Time Sampling.Proceedings of the AAAI Conference on Artificial Intelligence35, 5 (2021), 4582–4589. doi:10.1609...

    2021 doi

  20. [28]

    Xin Zhang, Weiliang Li, Rui Li, Zihang Fu, Tongyi Tang, Zhengyu Zhang, Wen-Yen Chen, Nima Noorshams, Nirav Jasapara, Xiaowen Ding, Ellie Wen, and Xue Feng. 2025. Personalized Interpolation: Achieving Efficient Conver- sion Estimation with Flexible Optimization Windows. InProce...

    2025

  21. [29]

    Han Zhu, Junqi Jin, Chang Tan, Fei Pan, Yifan Zeng, Han Li, and Kun Gai. 2017. Optimized Cost per Click in Taobao Display Advertising. InProceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD ’17). Association for Computing Mach...

    2017

  22. [30]

    Yuya Yoshikawa and Yusaku Imai. 2020. A Nonparametric Delayed Feed- back Model for Conversion Rate Prediction. InProceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD ’20). Association for Computing Machinery, New York, NY, USA, ...

    2020

  23. [33]

    The delay head uses separate embeddings for the fields in𝐺, a two-layer ReLU MLP with widths[64, 32], and a𝐾-way softmax; it shares no trainable parameters with the CVR branch

    For 𝑣= 30days, the 𝐾= 12delay intervals use boundaries 0, 1, 2, 4, 8, 12, 24hours followed by2 , 4, 7, 14, 21, and30days. The delay head uses separate embeddings for the fields in𝐺, a two-layer ReLU MLP with widths[64, 32], and a𝐾-way softmax; it shares no trainable parameters...

  24. [195]

    doi:10.1145/3298689.3347002

  25. [2022]

    InAdvances in Neural Information Processing Systems, S

    Incrementality Bidding via Reinforcement Learning under Mixed and Delayed Rewards. InAdvances in Neural Information Processing Systems, S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh (Eds.), Vol. 35. Curran Associates, Inc., Red Hook, NY, USA, 2142–2153. doi...

    doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.