Pith. sign in

REVIEW 4 major objections 9 minor 49 references

AliBoost: Ecological Boosting Framework in Alibaba Platform

T0 review · 4 major / 9 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read AliBoost claims that adding a staged, bidding-driven boosting layer to a billion-scale recommender system can systematically rescue cold items from popularity bias, reporting over a billion items cold-started and cold-item clicks and GMV…

desk verdict Genuine industrial artifact, but the +60% cold-start claim conflates forced boosting exposure with organic growth. read the letter →

arxiv 2506.00954 v1 pith:NA3LYEIF submitted 2025-06-01 cs.IR

Qijie Shen , Yuanchen Bei , Zihong Huang , Jialin Zhu , Keqin Xu , Boya Du , Jiawei Tang , Yuning Jiang
show 3 more authors
Feiran Huang Xiao Huang Hao Chen
This is my paper · ORCID
classification cs.IR
keywords ecologicalboostingcold-startrecommendationitem-orientedbiddingCTRpredictionpopularitybiasMattheweffectbillion-scalerecommendersystemsindustrialdeployment
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

Online marketplaces naturally reward popular items, so newly listed products often starve for attention before they get a chance to prove themselves. This paper claims that Alibaba's AliBoost framework fixes that by adding a separate 'boosting' layer that hands new items staged, performance-gated extra exposure. The layer is driven by a specialized click-rate predictor for cold items and a bidding rule that sends each item to users most likely to engage with it. The reported result is that, over six months across Alibaba's main platforms, over a billion new items were cold-started, cold-item clicks and GMV rose by more than 60% within 180 days, and the share of items failing to reach 10 daily page views fell from 41.1% to 24.5%.

What carries the argument

The load-bearing object is the item-oriented bidding rule, anchored by a single potential estimate. For a user-item pair $(u,i)$ at time $t$, the item's bid is the cold-start CTR prediction $\hat{y}^{cold}_{u,i}$, and the user's ideal price is $P_{40,i}\cdot S_{i,t}\cdot U_u$, where $P_{40,i}$ is the 40th percentile of the item's predicted CTR distribution, $S_{i,t}$ is a boosting-speed factor that raises or lowers the price to keep delivery at a target rate, and $U_u$ is a user-preference factor penalizing fatigue and rewarding activity. The delivery rule is simply $\mathrm{Deliver}_{u,i,t} = 1$ if $\mathrm{Bid}_{u,i} > \mathrm{Price}_{u,i,t}$, otherwise 0. This single comparison converts an abstract 'item potential' into a concrete, controllable exposure budget: $P_{40}$ sets the base threshold, the speed factor turns the remaining budget into a time schedule, and the user factor steers each impression to the most receptive users. The same $P_{40}$ rank also assigns the item to its boosting stage (stage 1 below the 70th percentile rank, stage 2 to the 90th, stage 3 above it), so one calibrated percentile drives both the tier structure and the auction threshold.

What would settle it

Calibration audit: use a held-out set of cold items that completed their first 180 days, and compare each item's launch-time $P_{40}$ against its observed 180-day clicks and GMV. If the observed click rate is flat or decreasing across $P_{40}$ bins, or if items assigned to Stage 3 do not outperform Stage 1 items in final engagement, then the potential signal that gates the whole framework is not what is carrying the reported gains. A second, sharper check: rerun the item A/B test with $P_{40}$ replaced by a random percentile; if the 40% reduction in neglected items persists, the predictor is not the driver.

Watch

Extended reading notes

Core claim

The paper's central claim is that the rich-get-richer bias of natural recommendation can be counteracted at billion-user scale by an explicit item-oriented boosting layer. AliBoost first predicts each cold item's click-through rate with a stacking fine-tuned model built on the platform's foundation CTR model, then summarizes the item's potential as the 40th percentile ($P_{40}$) of its predicted CTR distribution over a large user sample. That single number anchors two decisions: which boosting stage the item enters, and the base 'ideal price' in a bidding rule where an item is delivered to a user only if the item's predicted CTR exceeds the user's ideal price, adjusted by a boosting-speed factor and a user-preference (activity and fatigue) factor. Promotion and exit rules compare each item's boosted CTR against category-average CTR thresholds, so well-performing items ascend through larger exposure tiers while underperformers are phased out. Deployed across Alibaba's mainstream platforms, the framework cold-started over a billion items and, within 180 days, lifted cold-item clicks by 76% and GMV by 72%, while cutting the fraction of items with fewer than 10 daily page views from 41.1% to 24.5%.

Load-bearing premise

The entire allocation rests on the cold-start CTR model's predicted click-rate distribution being an honest measure of each new item's potential; if $P_{40}$ is biased for cold items, the framework will systematically give exposure to the wrong items.

Editorial extensions

If this is right

  • Items that clear the promotion threshold $\gamma^{(k)}\cdot \mathrm{CTR}^{category}_i$ move into larger exposure budgets, so a successful cold item compounds its visibility rather than plateauing.
  • The exit rule prunes underperformers, so boosting does not simply flood the feed with junk; the Non-Disturbance Principle requires expected boosted CTR to stay at or above 1.2 times the natural recommendation CTR.
  • Top-item dominance weakens: the share of daily top-100 items that stay in the top 100 after 7 days falls by up to 71.6%, meaning exposure is redistributed toward newer items.
  • Platform-wide totals still move: overall PV, clicks, payments, and GMV rise 2.0 to 4.7%, so the ecological fix does not sacrifice headline business metrics.
  • The framework operates in a live cold-item channel at up to 120,000 QPS with sub-20ms response, showing the mechanism is deployable under production latency budgets.

Reading between the lines

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

  • Editorial inference: the use of one calibrated percentile as both tier-gate and auction base price makes AliBoost a template rather than a bespoke system; any platform with a predicted-value distribution for fresh items could adopt the same staged bidding structure with almost no parameter tuning.
  • Editorial inference: the 40% drop in items failing to reach 10 daily page views reframes cold-start success from 'high accuracy on day one' to 'survival through a staged incubation'; that reframing could be a more useful KPI for marketplace health than standard offline metrics.
  • Editorial inference: because $P_{40}$ is a quantile of a model's output, the framework inherits every bias of that model; coupling it with uncertainty quantification (e.g., using the width of the predicted CTR distribution for items with very few interactions) is a natural extension the paper does not explore.
  • Editorial inference: the bidding mechanism effectively makes the platform the price-setter for item attention, which connects to market-design questions about how to allocate a scarce resource (user attention) when the 'advertiser' is the platform itself rather than a paying customer.
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

4 major / 9 minor

Summary. This paper describes AliBoost, an industrial 'ecological boosting' framework deployed on Alibaba's Taobao platform to accelerate the growth of cold-start items. The framework has three main components: (i) a tiered boosting structure with stage budgets, promotion rules, and exit rules (Eqs. 4-8); (ii) a Stacking Fine-Tuning Cold Predictor that augments a foundation CTR model with cold-item features and uses the 40th percentile (P40) of the predicted CTR distribution over a sampled user set as an item-potential measure (Eqs. 9-16); and (iii) an Item-Oriented Bidding Boosting mechanism in which an item is delivered to a user when the model-predicted CTR (the bid) exceeds a user-item ideal price that combines P40 with a boosting-speed factor and a user-preference factor (Eqs. 17-22). The paper claims six months of deployment, over one billion cold-started items, a reduction from 41.1% to 24.5% in the share of cold items failing to reach 10 daily PV, and 60-76% lifts in PV, CLICK, PAY, and GMV for 180-day cold-item cohorts (Table 2). Evidence includes platform-level A/B comparisons, ablations of the boosting and bidding components, offline AUC comparisons (0.54-0.64) against DeepFM/DIN baselines with ALDI and ColdLLM, and category-level analyses.

Significance. If the online results are internally valid, this is a valuable industrial case study: the deployment is real and large-scale, the design rationale is explicit (the direct-boosting versus amplified-natural-exposure decomposition in Eq. (1) is a useful organizing concept), the operational details are concrete (120k QPS, average latency under 20 ms, 15-minute model and price update cycles), and the ablation coverage (Tables 3, 5, and the category-level Table 8) is more extensive than is typical for an industrial paper. The framework also makes falsifiable predictions: natural-only growth metrics and exit-inclusive cohort outcomes would either confirm or refute the central claim. Where the paper falls short is evidence quality around that central claim: the reported lifts do not separate forced boosting exposure from organic growth, the long-horizon cohort may be a survivor set, and the potential measure P40 is built on a near-random CTR model (Table 7) without any calibration or validation. The framework is credible as an engineering contribution, but the headline '60% within 180 days' result is not established as stated.

major comments (4)
  1. [§4.2, Table 2; Eqs. (1), (25)] The central quantitative claim—clicks and GMV of cold items increase by over 60% within 180 days—is not supported by the metrics as reported, because Table 2 does not separate direct boosting exposure from organic growth. The treatment arm injects exposure by construction (Eq. (17), plus the budget allocations of Eqs. (4)-(8)), so boosted items accumulate PV, CLICK, PAY, and GMV from forced deliveries regardless of item quality. The paper itself defines the direct/amplified decomposition in Eq. (1) and even defines ROI as natural cold-item PV divided by boost PV in Eq. (25), yet none of the online tables report natural-only PV, natural CTR, or organic GMV per exposure for the 180-day cohort, nor the absolute exposure volumes in treatment versus control. Please report natural-only and boosted-only decompositions for a fixed cohort, with exposure-normalized metrics and confidence intervals (or establish that exposure volume is matched between arms); absent that, the 'successfully cold-starting over a billion items' claim may conflate mechanical exposure-driven gains with genuine organic growth.
  2. [§3.2, Eq. (7); §4.2.2; §4.1.3] The definition of the 180-day cohort is ambiguous with respect to the exit rule in Eq. (7): items whose boosting CTR falls below gamma(k)*CTR_category leave the boosting pipeline, and the control bucket of the item A/B test (Section 4.1.3) has no analogous removal rule. If the 'Boosted 180 Days' rows of Table 2 are computed only over items that ever passed through the pipeline or that remain in it, the reported lifts can be inflated by removing the worst-performing cold items from the treated arm after their initial poor performance. Specify whether each 'Boosted X Days' row is a fixed cohort defined at launch, state explicitly whether exited items remain in the denominator, and report intent-to-treat metrics for all launched items, including those that exited and their post-exit outcomes.
  3. [§3.3.3, §3.4.1, Table 7] The potential measure P40 (Eqs. (14)-(16)), the stage assignment, the bidding price Bid=y_cold (Eq. (17)), and the base ideal price in Eq. (18) all derive from the cold-start CTR model, but the offline AUC of that model ranges only from 0.54 to 0.64 (Table 7), barely above the 0.5 random baseline. Low AUC does not automatically invalidate the framework, since P40 is used only as a relative signal, but the paper provides no evidence that the predicted CTR distribution is informative for potential: there is no calibration analysis and no check that P40 at launch correlates with later organic CTR, retention, or GMV. Please add confidence intervals for the AUC gains, a calibration curve, and a rank-correlation test between launch-time P40 and 7/30/180-day organic outcomes. In addition, the construction of the user sample used to compute D_i in Eq. (14) is unspecified (sample size N and sampling procedure), which makes the potential measure impossible to reproduce.
  4. [§4.1.3, Table 2, Figure 4] The statistical reporting of the online results is currently insufficient to assess their reliability. No confidence intervals, significance tests, or day-level variation are given for any online metric in Tables 2-5 or Figure 4; the A/B design in Section 4.1.3 does not specify the number of hash buckets, the treatment/control allocation ratio, or whether the reported percentages are per-exposure or aggregate; and the 'reverse experimental setting' used as the comparison in Section 4.2.1 is never described. At minimum, state the allocation mechanism, the unit of analysis (user vs. item), and the uncertainty of the headline 60-76% numbers (for example, bootstrap intervals over days).
minor comments (9)
  1. [§3.3.3] The sentence defining P40 ('captures the CTR threshold that 40% of the sampled users are predicted to exceed') is inconsistent with the standard definition of the 40th percentile, below which 40% of the values fall; the described quantity is the 60th percentile. Please fix the wording or the intended percentile.
  2. [Table 1] The column header 'Catagories' is misspelled, and the category counts are not explained (24,568 for the foundation training data versus 9,567 for the cold-start data, despite the cold-start data containing five times more items).
  3. [Tables 2 and 8] The 'ALL' row of Table 8 (7-Day PV 31%, 30-Day PV 45%) does not match the corresponding rows of Table 2 (Boosted 7 Days +29.20% PV; Boosted 30 Days +44.93% PV); reconcile the numbers or state the different measurement windows.
  4. [§4.4, Table 4] PCTR is reported (+10.71%, +11.40%) but is not defined in the metrics list of §4.1.2, and the row 'Offline AUC +2.24%' gives only a relative gain with no absolute AUC value and no link to the absolute AUCs in Table 7.
  5. [§4.2.1] The 'reverse experimental setting' used as the baseline for the platform overall comparison in Table 2 is not described; specify the control condition (for example, no boosting channel) and whether user- or item-level hashing was used for this comparison.
  6. [§4.3.2, Table 3] The 'Fix Boosting' row of Table 3 contains only dashes; state that it is the reference (0%) level so that the percentage gains in neighboring rows are interpretable.
  7. [Appendix A] The threshold-adjustment enumeration reads 'two key factors' followed by items numbered (1) and (3) with no item (2); fix the numbering and the count.
  8. [§3.2 and §3.3.3] The relationship between the two stage-assignment mechanisms is unclear: Eq. (16) assigns an initial stage from the P40 rank (allowing a newly launched item to start directly at Stage 3), whereas Eq. (7) promotes or exits items stage by stage, and Eq. (4) states that an item at stage k enjoys cumulative budgets from all preceding stages. Clarify how an item can start at a high stage without passing through lower stages, and how the cumulative budget statement is reconciled with the rank-based initial assignment.
  9. [Abstract and §1] The claim of 'successfully cold-starting over a billion new items' is not defined operationally; specify the success criterion (for example, reaching a minimum daily PV, exiting the boosting pipeline, or achieving a target CTR).

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central online-lift result is an empirical A/B measurement rather than an algebraic consequence of the framework equations, and the self-citations are not load-bearing.

full rationale

The paper's central empirical claim (a greater-than-60% increase in cold-item clicks and GMV within 180 days, and over one billion cold-started items) is presented as an online A/B comparison against a reverse experimental setting in Table 2. That outcome is an external measurement, not derived from the framework's equations, so it cannot be an algebraic consequence of the model. The potential metric P40 is introduced as the 40th percentile of the cold-start CTR model's predicted distribution (Eq. 14) and is used both to assign boosting stages (Eq. 16) and as the base price in the auction rule (Eq. 18), with the item's bid set to the same model's prediction (Eq. 17). This is a control-loop design choice rather than a circular derivation: the paper does not claim that P40 predicts some independently measured 'true potential' and then use that same claim as the evidence of success. The self-citations (e.g., ALDI [13] and ColdLLM [12] used as offline baselines) are not load-bearing; these baselines are evaluated in this paper and are not invoked as a uniqueness theorem or as the justification for the framework's design. The possible survivorship and traffic-source confounds in Table 2 are experimental-validity concerns, but the paper does not define the 180-day improvement as equal to forced boosting exposure by any quoted equation, so under the reduction-based standard they do not constitute circularity. The low offline AUC values are a correctness risk, not a circularity indicator.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The framework's design depends on multiple hand-picked thresholds and heuristic formulas. The key free parameters are the safety factors gamma, the rank thresholds for stage assignment, the 40th percentile potential measure, the speed factor coefficients, loss weights, stage budgets, and the user preference factor formula. The axioms are unmeasured assumptions about how boosting CTR translates to organic exposure, how to define non-disturbance, and which benchmark should drive promotion/exit decisions. None of these are derived from first principles or validated with controlled experiments.

free parameters (7)
  • Safety factor gamma and stage thresholds gamma(k) = gamma = 1.2 (typical), gamma(k) increasing with stage
    Used in non-disturbance principle (Eq. 3) and promotion/exit rules (Eqs. 5-7); values are chosen by hand, no sensitivity analysis.
  • Stage assignment rank thresholds = 70% and 90%
    Items with P40 rank below 70% get stage 1, 70-90% stage 2, above 90% stage 3 (Eq. 16). Chosen without empirical tuning or sensitivity analysis.
  • Potential percentile P40 = 40th percentile
    The 40th percentile of the predicted CTR distribution is used as the potential score and as the base price in bidding (Eqs. 14-18). The choice of 40% is arbitrary.
  • Boosting speed factor hyperparameters delta_p, delta_q, delta_d = not reported
    Weights for current and past speed errors in Eq. (21); exact values are withheld.
  • Loss weights omega_s and regularization alpha = not reported
    Weighted loss across data sources and regularization strength in Eq. (12); values not given.
  • Stage exposure budgets B_i^(k) and target speeds V_target = not reported
    The tiered budget sizes and target delivery speeds determine how much exposure each item receives; not specified.
  • User preference factor constants = 10 activity grades, functional form ln(U_tired) * U_active^{-1/2}
    The user activity grading and the exact formula in Eq. (22) are heuristic choices without validation or sensitivity analysis.
assumptions (5)
  • domain assumption The amplification function alpha(CTR_boost) in Eq. (1) is monotonically increasing, with exponential growth for high CTR.
    No empirical measurement or derivation is given; it underlies the claim that boosting produces amplified organic exposure.
  • domain assumption Maintaining expected boosting CTR above gamma times natural CTR is sufficient to avoid disturbing natural recommendations.
    Non-disturbance principle Eq. (3); the threshold and mechanism are asserted, not tested.
  • domain assumption Category-mean CTR is the correct reference benchmark for promotion and exit decisions.
    Promotion and exit rules in Eqs. (5)-(7) compare item CTR to category average; no evidence that this benchmark optimizes ecosystem health.
  • ad hoc to paper The user preference factor formula U_u = ln(U_tired) * U_active^{-1/2} captures user fatigue and activity effects.
    Eq. (22) is presented without derivation or empirical calibration; the functional form appears chosen for convenience.
  • domain assumption The deterministic bidding rule (Eq. 17) with a static threshold yields an optimal or near-optimal delivery policy under budget and speed constraints.
    No optimization derivation is provided; the rule is a heuristic threshold comparison.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AliBoost: Ecological Boosting Framework in Alibaba Platform." pith.science (2026). https://pith.science/paper/NA3LYEIF

@misc{pith2026250600954,
  author       = {Pith},
  title        = {Pith review of: AliBoost: Ecological Boosting Framework in Alibaba Platform},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NA3LYEIF}},
  note         = {Machine review of arXiv:2506.00954}
}
read the original abstract

Maintaining a healthy ecosystem in billion-scale online platforms is challenging, as users naturally gravitate toward popular items, leaving cold and less-explored items behind. This ''rich-get-richer'' phenomenon hinders the growth of potentially valuable cold items and harms the platform's ecosystem. Existing cold-start models primarily focus on improving initial recommendation performance for cold items but fail to address users' natural preference for popular content. In this paper, we introduce AliBoost, Alibaba's ecological boosting framework, designed to complement user-oriented natural recommendations and foster a healthier ecosystem. AliBoost incorporates a tiered boosting structure and boosting principles to ensure high-potential items quickly gain exposure while minimizing disruption to low-potential items. To achieve this, we propose the Stacking Fine-Tuning Cold Predictor to enhance the foundation CTR model's performance on cold items for accurate CTR and potential prediction. AliBoost then employs an Item-oriented Bidding Boosting mechanism to deliver cold items to the most suitable users while balancing boosting speed with user-personalized preferences. Over the past six months, AliBoost has been deployed across Alibaba's mainstream platforms, successfully cold-starting over a billion new items and increasing both clicks and GMV of cold items by over 60% within 180 days. Extensive online analysis and A/B testing demonstrate the effectiveness of AliBoost in addressing ecological challenges, offering new insights into the design of billion-scale recommender systems.

Figures

Figures reproduced from arXiv: 2506.00954 by the authors.

Figure 1
Figure 1. Distribution comparison of daily page views (PV) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Toy example of three types of recommendation [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overall architecture of the AliBoost framework. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Relative changes in recommendation metrics for [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The deployment architecture of AliBoost. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 25 canonical work pages

  1. [1]

    Carlos Carrion, Zenan Wang, Harikesh Nair, Xianghong Luo, Yulin Lei, Peiqin Gu, Xiliang Lin, Wenlong Chen, Junsheng Jin, Fanan Zhu, et al. 2023. Blending advertising with organic content in e-commerce via virtual bids. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 15476–15484

    work page 2023

  2. [2]

    Hao Chen, Yuanchen Bei, Qijie Shen, Yue Xu, Sheng Zhou, Wenbing Huang, Feiran Huang, Senzhang Wang, and Xiao Huang. 2024. Macro graph neural networks for online billion-scale recommender systems. InProceedings of the ACM web conference 2024. 3598–3608

    2024

  3. [3]

    Jiawei Chen, Hande Dong, Xiang Wang, Fuli Feng, Meng Wang, and Xiangnan He. 2023. Bias and debias in recommender system: A survey and future directions. ACM Transactions on Information Systems41, 3 (2023), 1–39

    2023

  4. [4]

    Li Chen, Yonghua Yang, Ningxia Wang, Keping Yang, and Quan Yuan. 2019. How serendipity improves user satisfaction with recommendations? a large-scale user evaluation. InThe world wide web conference. 240–250

    work page 2019

  5. [5]

    Giovanni Luca Ciampaglia, Azadeh Nematzadeh, Filippo Menczer, and Alessandro Flammini. 2018. How algorithmic popularity bias hinders or promotes quality. Scientific reports8, 1 (2018), 15951

    work page 2018

  6. [6]

    Francesco Fabbri, Maria Luisa Croci, Francesco Bonchi, and Carlos Castillo. 2022. Exposure inequality in people recommender systems: the long-term effects. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 16. 194–204

    work page 2022

  7. [7]

    Chongming Gao, Kexin Huang, Jiawei Chen, Yuan Zhang, Biao Li, Peng Jiang, Shiqi Wang, Zhong Zhang, and Xiangnan He. 2023. Alleviating matthew effect of offline reinforcement learning in interactive recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 238–248

    work page 2023

  8. [8]

    Chen Gao, Yu Zheng, Nian Li, Yinfeng Li, Yingrong Qin, Jinghua Piao, Yuhan Quan, Jianxin Chang, Depeng Jin, Xiangnan He, et al. 2023. A survey of graph neural networks for recommender systems: Challenges, methods, and directions. ACM Transactions on Recommender Systems1, 1 (2023), 1–51

    2023

Show all 49 references
  1. [9]

    Fabrizio Germano, Vicenç Gómez, and Gaël Le Mens. 2019. The few-get-richer: a surprising consequence of popularity-based rankings?. InThe World Wide Web Conference. 2764–2770

    2019

  2. [10]

    Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. 2017. DeepFM: a factorization-machine based neural network for CTR prediction. In Proceedings of the 26th International Joint Conference on Artificial Intelligence. 1725–1731

    2017

  3. [11]

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. 2020. Lightgcn: Simplifying and powering graph convolution network for recommendation. InProceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval. 639–648

    2020

  4. [12]

    Feiran Huang, Yuanchen Bei, Zhenghang Yang, Junyi Jiang, Hao Chen, Qijie Shen, Senzhang Wang, Fakhri Karray, and Philip S Yu. 2025. Large Language Model Simulator for Cold-Start Recommendation. InProceedings of the Eighteenth ACM International Conference on Web Search and Data...

    2025

  5. [13]

    Feiran Huang, Zefan Wang, Xiao Huang, Yufeng Qian, Zhetao Li, and Hao Chen

  6. [14]

    Geng Ji, Wentao Jiang, Jiang Li, Fahmid Morshed Fahid, Zhengxing Chen, Yinghua Li, Jun Xiao, Chongxi Bao, and Zheqing Zhu. 2024. Learning to bid and rank together in recommendation systems.Machine Learning113, 5 (2024), 2559–2573

    2024

  7. [15]

    Anastasiia Klimashevskaia, Mehdi Elahi, Dietmar Jannach, Lars Skjærven, Astrid Tessem, and Christoph Trattner. 2023. Evaluating The Effects of Calibrated Pop- ularity Bias Mitigation: A Field Study. InProceedings of the 17th ACM Conference on Recommender Systems. 1084–1089

    2023

  8. [16]

    Anastasiia Klimashevskaia, Dietmar Jannach, Mehdi Elahi, and Christoph Trat- tner. 2024. A survey on popularity bias in recommender systems.User Modeling and User-Adapted Interaction34, 5 (2024), 1777–1834

    2024

  9. [17]

    Ioannis Konstas, Vassilios Stathopoulos, and Joemon M Jose. 2009. On social networks and collaborative recommendation. InProceedings of the 32nd interna- tional ACM SIGIR conference on Research and development in information retrieval. 195–202

    2009

  10. [18]

    Hoyeop Lee, Jinbae Im, Seongwon Jang, Hyunsouk Cho, and Sehee Chung. 2019. Melu: Meta-learned user preference estimator for cold-start recommendation. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 1073–1082

    2019

  11. [19]

    Yunqi Li, Hanxiong Chen, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2021. User-oriented fairness in recommendation. InProceedings of the web conference

    2021

  12. [20]

    Ruochen Liu, Hao Chen, Yuanchen Bei, Qijie Shen, Fangwei Zhong, Senzhang Wang, and Jianxin Wang. 2024. Fine Tuning Out-of-Vocabulary Item Recommen- dation with User Sequence Imagination. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems

    2024

  13. [21]

    Siyi Liu and Yujia Zheng. 2020. Long-tail session-based recommendation. In Proceedings of the 14th ACM conference on recommender systems. 509–514

    2020

  14. [22]

    Taichi Liu, Chen Gao, Zhenyu Wang, Dong Li, Jianye Hao, Depeng Jin, and Yong Li. 2023. Uncertainty-aware Consistency Learning for Cold-Start Item Recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. ...

    2023

  15. [23]

    Mehrbakhsh Nilashi, Othman Ibrahim, and Karamollah Bagherifard. 2018. A recommender system based on collaborative filtering using ontology and di- mensionality reduction techniques.Expert Systems with Applications92 (2018), 507–520

    2018

  16. [24]

    Weitong Ou, Bo Chen, Xinyi Dai, Weinan Zhang, Weiwen Liu, Ruiming Tang, and Yong Yu. 2023. A survey on bid optimization in real-time bidding display advertising.ACM Transactions on Knowledge Discovery from Data18, 3 (2023), 1–31

    2023

  17. [25]

    Matjaž Perc. 2014. The Matthew effect in empirical data.Journal of The Royal Society Interface11, 98 (2014), 20140378

    2014

  18. [26]

    Shuai Wang, Kun Zhang, Le Wu, Haiping Ma, Richang Hong, and Meng Wang

  19. [27]

    Wenjie Wang, Fuli Feng, Xiangnan He, Xiang Wang, and Tat-Seng Chua. 2021. Deconfounded recommendation for alleviating bias amplification. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 1717– 1725

    2021

  20. [28]

    Tianxin Wei, Fuli Feng, Jiawei Chen, Ziwei Wu, Jinfeng Yi, and Xiangnan He

  21. [29]

    Yinwei Wei, Xiang Wang, Qi Li, Liqiang Nie, Yan Li, Xuanping Li, and Tat-Seng Chua. 2021. Contrastive learning for cold-start recommendation. InProceedings of the 29th ACM International Conference on Multimedia. 5382–5390

    2021

  22. [30]

    Hongyi Wen, Xinyang Yi, Tiansheng Yao, Jiaxi Tang, Lichan Hong, and Ed H Chi

  23. [31]

    InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining

    Model-agnostic counterfactual reasoning for eliminating popularity bias in recommender system. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 1791–1800

  24. [32]

    Hao Zhang, Gang Chen, Beng Chin Ooi, Kian-Lee Tan, and Meihui Zhang. 2015. In-memory big data management and processing: A survey.IEEE Transactions on Knowledge and Data Engineering27, 7 (2015), 1920–1948

    2015

  25. [33]

    Haoqi Zhang, Lvyin Niu, Zhenzhe Zheng, Zhilin Zhang, Shan Gu, Fan Wu, Chuan Yu, Jian Xu, Guihai Chen, and Bo Zheng. 2023. A personalized automated bidding framework for fairness-aware online advertising. InProceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery an...

    2023

  26. [34]

    Junsan Zhang, Sini Wu, Te Wang, Fengmei Ding, and Jie Zhu. 2025. Relieving popularity bias in recommendation via debiasing representation enhancement. Complex & Intelligent Systems11, 1 (2025), 34

    2025

  27. [35]

    Yuhao Yang, Chao Huang, Lianghao Xia, Chunzhen Huang, Da Luo, and Kangyi Lin. 2023. Debiased contrastive learning for sequential recommendation. In Proceedings of the ACM web conference 2023. 1063–1073

    2023

  28. [36]

    Weizhi Zhang, Yuanchen Bei, Liangwei Yang, Henry Peng Zou, Peilin Zhou, Aiwei Liu, Yinghui Li, Hao Chen, Jianling Wang, Yu Wang, et al . 2025. Cold- Start Recommendation towards the Era of Large Language Models (LLMs): A Comprehensive Survey and Roadmap.arXiv preprint arXiv:25...

    2025 arXiv

  29. [37]

    Mark Zhao, Niket Agarwal, Aarti Basant, Buğra Gedik, Satadru Pan, Mustafa Ozdal, Rakesh Komuravelli, Jerry Pan, Tianshu Bao, Haowei Lu, et al . 2022. Understanding data storage and ingestion for large-scale deep recommendation model training: Industrial product. InProceedings ...

    2022

  30. [38]

    Chang Zhou, Jianxin Ma, Jianwei Zhang, Jingren Zhou, and Hongxia Yang. 2021. Contrastive learning for debiased candidate generation in large-scale recom- mender systems. InProceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 3985–3995

    2021

  31. [39]

    Jiahao Zhang, Rui Xue, Wenqi Fan, Xin Xu, Qing Li, Jian Pei, and Xiaorui Liu

  32. [40]

    Zhihui Zhou, Lilin Zhang, and Ning Yang. 2023. Contrastive collaborative filtering for cold-start item recommendation. InProceedings of the ACM Web Conference

    2023

  33. [41]

    Yongchun Zhu, Ruobing Xie, Fuzhen Zhuang, Kaikai Ge, Ying Sun, Xu Zhang, Leyu Lin, and Juan Cao. 2021. Learning to warm up cold item embeddings for cold- start recommendation with meta scaling and shifting networks. InProceedings of the 44th International ACM SIGIR Conference ...

    2021

  34. [42]

    Ziwei Zhu, Yun He, Xing Zhao, and James Caverlee. 2021. Popularity bias in dynamic recommendation. InProceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining. 2439–2449

    2021

  35. [43]

    Ziwei Zhu, Yun He, Xing Zhao, Yin Zhang, Jianling Wang, and James Caverlee

  36. [44]

    Guorui Zhou, Xiaoqiang Zhu, Chenru 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. InProceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining. ...

    2018

  37. [49]

    InProceedings of the 14th ACM International Conference on Web Search and Data Mining

    Popularity-opportunity bias in collaborative filtering. InProceedings of the 14th ACM International Conference on Web Search and Data Mining. 85–93. AliBoost: Ecological Boosting Framework in Alibaba Platform KDD ’25, August 3–7, 2025, Toronto, ON, Canada. Cold Item Pool User ...

    2025

  38. [2021]

    InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Privileged graph distillation for cold start recommendation. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1187–1196

  39. [2022]

    InProceedings of the ACM Web Conference 2022

    Distributionally-robust recommendations for improving worst-case user experience. InProceedings of the ACM Web Conference 2022. 3606–3610

    2022

  40. [2023]

    InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Aligning distillation for cold-start item recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1147–1157

  41. [2024]

    In Proceedings of the ACM on Web Conference 2024

    Linear-Time Graph Neural Networks for Scalable Recommendations. In Proceedings of the ACM on Web Conference 2024. 3533–3544

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.